fs/btrfs/free-space-cache.c

   1 /*
   2  * Copyright (C) 2008 Red Hat.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/pagemap.h>
  20 #include <linux/sched.h>
  21 #include <linux/slab.h>
  22 #include <linux/math64.h>
  23 #include <linux/ratelimit.h>
  24 #include "ctree.h"
  25 #include "free-space-cache.h"
  26 #include "transaction.h"
  27 #include "disk-io.h"
  28 #include "extent_io.h"
  29 #include "inode-map.h"
  30 #include "volumes.h"
  31
  32 #define BITS_PER_BITMAP         (PAGE_SIZE * 8UL)
  33 #define MAX_CACHE_BYTES_PER_GIG SZ_32K
  34
  35 struct btrfs_trim_range {
  36         u64 start;
  37         u64 bytes;
  38         struct list_head list;
  39 };
  40
  41 static int link_free_space(struct btrfs_free_space_ctl *ctl,
  42                            struct btrfs_free_space *info);
  43 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
  44                               struct btrfs_free_space *info);
  45
  46 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
  47                                                struct btrfs_path *path,
  48                                                u64 offset)
  49 {
  50         struct btrfs_key key;
  51         struct btrfs_key location;
  52         struct btrfs_disk_key disk_key;
  53         struct btrfs_free_space_header *header;
  54         struct extent_buffer *leaf;
  55         struct inode *inode = NULL;
  56         int ret;
  57
  58         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
  59         key.offset = offset;
  60         key.type = 0;
  61
  62         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  63         if (ret < 0)
  64                 return ERR_PTR(ret);
  65         if (ret > 0) {
  66                 btrfs_release_path(path);
  67                 return ERR_PTR(-ENOENT);
  68         }
  69
  70         leaf = path->nodes[0];
  71         header = btrfs_item_ptr(leaf, path->slots[0],
  72                                 struct btrfs_free_space_header);
  73         btrfs_free_space_key(leaf, header, &disk_key);
  74         btrfs_disk_key_to_cpu(&location, &disk_key);
  75         btrfs_release_path(path);
  76
  77         inode = btrfs_iget(root->fs_info->sb, &location, root, NULL);
  78         if (IS_ERR(inode))
  79                 return inode;
  80         if (is_bad_inode(inode)) {
  81                 iput(inode);
  82                 return ERR_PTR(-ENOENT);
  83         }
  84
  85         mapping_set_gfp_mask(inode->i_mapping,
  86                         mapping_gfp_constraint(inode->i_mapping,
  87                         ~(__GFP_FS | __GFP_HIGHMEM)));
  88
  89         return inode;
  90 }
  91
  92 struct inode *lookup_free_space_inode(struct btrfs_root *root,
  93                                       struct btrfs_block_group_cache
  94                                       *block_group, struct btrfs_path *path)
  95 {
  96         struct inode *inode = NULL;
  97         u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
  98
  99         spin_lock(&block_group->lock);
 100         if (block_group->inode)
 101                 inode = igrab(block_group->inode);
 102         spin_unlock(&block_group->lock);
 103         if (inode)
 104                 return inode;
 105
 106         inode = __lookup_free_space_inode(root, path,
 107                                           block_group->key.objectid);
 108         if (IS_ERR(inode))
 109                 return inode;
 110
 111         spin_lock(&block_group->lock);
 112         if (!((BTRFS_I(inode)->flags & flags) == flags)) {
 113                 btrfs_info(root->fs_info,
 114                         "Old style space inode found, converting.");
 115                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
 116                         BTRFS_INODE_NODATACOW;
 117                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 118         }
 119
 120         if (!block_group->iref) {
 121                 block_group->inode = igrab(inode);
 122                 block_group->iref = 1;
 123         }
 124         spin_unlock(&block_group->lock);
 125
 126         return inode;
 127 }
 128
 129 static int __create_free_space_inode(struct btrfs_root *root,
 130                                      struct btrfs_trans_handle *trans,
 131                                      struct btrfs_path *path,
 132                                      u64 ino, u64 offset)
 133 {
 134         struct btrfs_key key;
 135         struct btrfs_disk_key disk_key;
 136         struct btrfs_free_space_header *header;
 137         struct btrfs_inode_item *inode_item;
 138         struct extent_buffer *leaf;
 139         u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC;
 140         int ret;
 141
 142         ret = btrfs_insert_empty_inode(trans, root, path, ino);
 143         if (ret)
 144                 return ret;
 145
 146         /* We inline crc's for the free disk space cache */
 147         if (ino != BTRFS_FREE_INO_OBJECTID)
 148                 flags |= BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 149
 150         leaf = path->nodes[0];
 151         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 152                                     struct btrfs_inode_item);
 153         btrfs_item_key(leaf, &disk_key, path->slots[0]);
 154         memzero_extent_buffer(leaf, (unsigned long)inode_item,
 155                              sizeof(*inode_item));
 156         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
 157         btrfs_set_inode_size(leaf, inode_item, 0);
 158         btrfs_set_inode_nbytes(leaf, inode_item, 0);
 159         btrfs_set_inode_uid(leaf, inode_item, 0);
 160         btrfs_set_inode_gid(leaf, inode_item, 0);
 161         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
 162         btrfs_set_inode_flags(leaf, inode_item, flags);
 163         btrfs_set_inode_nlink(leaf, inode_item, 1);
 164         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
 165         btrfs_set_inode_block_group(leaf, inode_item, offset);
 166         btrfs_mark_buffer_dirty(leaf);
 167         btrfs_release_path(path);
 168
 169         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 170         key.offset = offset;
 171         key.type = 0;
 172         ret = btrfs_insert_empty_item(trans, root, path, &key,
 173                                       sizeof(struct btrfs_free_space_header));
 174         if (ret < 0) {
 175                 btrfs_release_path(path);
 176                 return ret;
 177         }
 178
 179         leaf = path->nodes[0];
 180         header = btrfs_item_ptr(leaf, path->slots[0],
 181                                 struct btrfs_free_space_header);
 182         memzero_extent_buffer(leaf, (unsigned long)header, sizeof(*header));
 183         btrfs_set_free_space_key(leaf, header, &disk_key);
 184         btrfs_mark_buffer_dirty(leaf);
 185         btrfs_release_path(path);
 186
 187         return 0;
 188 }
 189
 190 int create_free_space_inode(struct btrfs_root *root,
 191                             struct btrfs_trans_handle *trans,
 192                             struct btrfs_block_group_cache *block_group,
 193                             struct btrfs_path *path)
 194 {
 195         int ret;
 196         u64 ino;
 197
 198         ret = btrfs_find_free_objectid(root, &ino);
 199         if (ret < 0)
 200                 return ret;
 201
 202         return __create_free_space_inode(root, trans, path, ino,
 203                                          block_group->key.objectid);
 204 }
 205
 206 int btrfs_check_trunc_cache_free_space(struct btrfs_root *root,
 207                                        struct btrfs_block_rsv *rsv)
 208 {
 209         u64 needed_bytes;
 210         int ret;
 211
 212         /* 1 for slack space, 1 for updating the inode */
 213         needed_bytes = btrfs_calc_trunc_metadata_size(root, 1) +
 214                 btrfs_calc_trans_metadata_size(root, 1);
 215
 216         spin_lock(&rsv->lock);
 217         if (rsv->reserved < needed_bytes)
 218                 ret = -ENOSPC;
 219         else
 220                 ret = 0;
 221         spin_unlock(&rsv->lock);
 222         return ret;
 223 }
 224
 225 int btrfs_truncate_free_space_cache(struct btrfs_root *root,
 226                                     struct btrfs_trans_handle *trans,
 227                                     struct btrfs_block_group_cache *block_group,
 228                                     struct inode *inode)
 229 {
 230         int ret = 0;
 231         struct btrfs_path *path = btrfs_alloc_path();
 232         bool locked = false;
 233
 234         if (!path) {
 235                 ret = -ENOMEM;
 236                 goto fail;
 237         }
 238
 239         if (block_group) {
 240                 locked = true;
 241                 mutex_lock(&trans->transaction->cache_write_mutex);
 242                 if (!list_empty(&block_group->io_list)) {
 243                         list_del_init(&block_group->io_list);
 244
 245                         btrfs_wait_cache_io(root, trans, block_group,
 246                                             &block_group->io_ctl, path,
 247                                             block_group->key.objectid);
 248                         btrfs_put_block_group(block_group);
 249                 }
 250
 251                 /*
 252                  * now that we've truncated the cache away, its no longer
 253                  * setup or written
 254                  */
 255                 spin_lock(&block_group->lock);
 256                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 257                 spin_unlock(&block_group->lock);
 258         }
 259         btrfs_free_path(path);
 260
 261         btrfs_i_size_write(inode, 0);
 262         truncate_pagecache(inode, 0);
 263
 264         /*
 265          * We don't need an orphan item because truncating the free space cache
 266          * will never be split across transactions.
 267          * We don't need to check for -EAGAIN because we're a free space
 268          * cache inode
 269          */
 270         ret = btrfs_truncate_inode_items(trans, root, inode,
 271                                          0, BTRFS_EXTENT_DATA_KEY);
 272         if (ret)
 273                 goto fail;
 274
 275         ret = btrfs_update_inode(trans, root, inode);
 276
 277 fail:
 278         if (locked)
 279                 mutex_unlock(&trans->transaction->cache_write_mutex);
 280         if (ret)
 281                 btrfs_abort_transaction(trans, ret);
 282
 283         return ret;
 284 }
 285
 286 static int readahead_cache(struct inode *inode)
 287 {
 288         struct file_ra_state *ra;
 289         unsigned long last_index;
 290
 291         ra = kzalloc(sizeof(*ra), GFP_NOFS);
 292         if (!ra)
 293                 return -ENOMEM;
 294
 295         file_ra_state_init(ra, inode->i_mapping);
 296         last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
 297
 298         page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
 299
 300         kfree(ra);
 301
 302         return 0;
 303 }
 304
 305 static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 306                        struct btrfs_root *root, int write)
 307 {
 308         int num_pages;
 309         int check_crcs = 0;
 310
 311         num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 312
 313         if (btrfs_ino(inode) != BTRFS_FREE_INO_OBJECTID)
 314                 check_crcs = 1;
 315
 316         /* Make sure we can fit our crcs into the first page */
 317         if (write && check_crcs &&
 318             (num_pages * sizeof(u32)) >= PAGE_SIZE)
 319                 return -ENOSPC;
 320
 321         memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
 322
 323         io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
 324         if (!io_ctl->pages)
 325                 return -ENOMEM;
 326
 327         io_ctl->num_pages = num_pages;
 328         io_ctl->root = root;
 329         io_ctl->check_crcs = check_crcs;
 330         io_ctl->inode = inode;
 331
 332         return 0;
 333 }
 334
 335 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
 336 {
 337         kfree(io_ctl->pages);
 338         io_ctl->pages = NULL;
 339 }
 340
 341 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
 342 {
 343         if (io_ctl->cur) {
 344                 io_ctl->cur = NULL;
 345                 io_ctl->orig = NULL;
 346         }
 347 }
 348
 349 static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
 350 {
 351         ASSERT(io_ctl->index < io_ctl->num_pages);
 352         io_ctl->page = io_ctl->pages[io_ctl->index++];
 353         io_ctl->cur = page_address(io_ctl->page);
 354         io_ctl->orig = io_ctl->cur;
 355         io_ctl->size = PAGE_SIZE;
 356         if (clear)
 357                 memset(io_ctl->cur, 0, PAGE_SIZE);
 358 }
 359
 360 static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
 361 {
 362         int i;
 363
 364         io_ctl_unmap_page(io_ctl);
 365
 366         for (i = 0; i < io_ctl->num_pages; i++) {
 367                 if (io_ctl->pages[i]) {
 368                         ClearPageChecked(io_ctl->pages[i]);
 369                         unlock_page(io_ctl->pages[i]);
 370                         put_page(io_ctl->pages[i]);
 371                 }
 372         }
 373 }
 374
 375 static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 376                                 int uptodate)
 377 {
 378         struct page *page;
 379         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
 380         int i;
 381
 382         for (i = 0; i < io_ctl->num_pages; i++) {
 383                 page = find_or_create_page(inode->i_mapping, i, mask);
 384                 if (!page) {
 385                         io_ctl_drop_pages(io_ctl);
 386                         return -ENOMEM;
 387                 }
 388                 io_ctl->pages[i] = page;
 389                 if (uptodate && !PageUptodate(page)) {
 390                         btrfs_readpage(NULL, page);
 391                         lock_page(page);
 392                         if (!PageUptodate(page)) {
 393                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 394                                            "error reading free space cache");
 395                                 io_ctl_drop_pages(io_ctl);
 396                                 return -EIO;
 397                         }
 398                 }
 399         }
 400
 401         for (i = 0; i < io_ctl->num_pages; i++) {
 402                 clear_page_dirty_for_io(io_ctl->pages[i]);
 403                 set_page_extent_mapped(io_ctl->pages[i]);
 404         }
 405
 406         return 0;
 407 }
 408
 409 static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 410 {
 411         __le64 *val;
 412
 413         io_ctl_map_page(io_ctl, 1);
 414
 415         /*
 416          * Skip the csum areas.  If we don't check crcs then we just have a
 417          * 64bit chunk at the front of the first page.
 418          */
 419         if (io_ctl->check_crcs) {
 420                 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
 421                 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 422         } else {
 423                 io_ctl->cur += sizeof(u64);
 424                 io_ctl->size -= sizeof(u64) * 2;
 425         }
 426
 427         val = io_ctl->cur;
 428         *val = cpu_to_le64(generation);
 429         io_ctl->cur += sizeof(u64);
 430 }
 431
 432 static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 433 {
 434         __le64 *gen;
 435
 436         /*
 437          * Skip the crc area.  If we don't check crcs then we just have a 64bit
 438          * chunk at the front of the first page.
 439          */
 440         if (io_ctl->check_crcs) {
 441                 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
 442                 io_ctl->size -= sizeof(u64) +
 443                         (sizeof(u32) * io_ctl->num_pages);
 444         } else {
 445                 io_ctl->cur += sizeof(u64);
 446                 io_ctl->size -= sizeof(u64) * 2;
 447         }
 448
 449         gen = io_ctl->cur;
 450         if (le64_to_cpu(*gen) != generation) {
 451                 btrfs_err_rl(io_ctl->root->fs_info,
 452                         "space cache generation (%llu) does not match inode (%llu)",
 453                                 *gen, generation);
 454                 io_ctl_unmap_page(io_ctl);
 455                 return -EIO;
 456         }
 457         io_ctl->cur += sizeof(u64);
 458         return 0;
 459 }
 460
 461 static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
 462 {
 463         u32 *tmp;
 464         u32 crc = ~(u32)0;
 465         unsigned offset = 0;
 466
 467         if (!io_ctl->check_crcs) {
 468                 io_ctl_unmap_page(io_ctl);
 469                 return;
 470         }
 471
 472         if (index == 0)
 473                 offset = sizeof(u32) * io_ctl->num_pages;
 474
 475         crc = btrfs_csum_data(io_ctl->orig + offset, crc,
 476                               PAGE_SIZE - offset);
 477         btrfs_csum_final(crc, (u8 *)&crc);
 478         io_ctl_unmap_page(io_ctl);
 479         tmp = page_address(io_ctl->pages[0]);
 480         tmp += index;
 481         *tmp = crc;
 482 }
 483
 484 static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
 485 {
 486         u32 *tmp, val;
 487         u32 crc = ~(u32)0;
 488         unsigned offset = 0;
 489
 490         if (!io_ctl->check_crcs) {
 491                 io_ctl_map_page(io_ctl, 0);
 492                 return 0;
 493         }
 494
 495         if (index == 0)
 496                 offset = sizeof(u32) * io_ctl->num_pages;
 497
 498         tmp = page_address(io_ctl->pages[0]);
 499         tmp += index;
 500         val = *tmp;
 501
 502         io_ctl_map_page(io_ctl, 0);
 503         crc = btrfs_csum_data(io_ctl->orig + offset, crc,
 504                               PAGE_SIZE - offset);
 505         btrfs_csum_final(crc, (u8 *)&crc);
 506         if (val != crc) {
 507                 btrfs_err_rl(io_ctl->root->fs_info,
 508                         "csum mismatch on free space cache");
 509                 io_ctl_unmap_page(io_ctl);
 510                 return -EIO;
 511         }
 512
 513         return 0;
 514 }
 515
 516 static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
 517                             void *bitmap)
 518 {
 519         struct btrfs_free_space_entry *entry;
 520
 521         if (!io_ctl->cur)
 522                 return -ENOSPC;
 523
 524         entry = io_ctl->cur;
 525         entry->offset = cpu_to_le64(offset);
 526         entry->bytes = cpu_to_le64(bytes);
 527         entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
 528                 BTRFS_FREE_SPACE_EXTENT;
 529         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 530         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 531
 532         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 533                 return 0;
 534
 535         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 536
 537         /* No more pages to map */
 538         if (io_ctl->index >= io_ctl->num_pages)
 539                 return 0;
 540
 541         /* map the next page */
 542         io_ctl_map_page(io_ctl, 1);
 543         return 0;
 544 }
 545
 546 static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
 547 {
 548         if (!io_ctl->cur)
 549                 return -ENOSPC;
 550
 551         /*
 552          * If we aren't at the start of the current page, unmap this one and
 553          * map the next one if there is any left.
 554          */
 555         if (io_ctl->cur != io_ctl->orig) {
 556                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 557                 if (io_ctl->index >= io_ctl->num_pages)
 558                         return -ENOSPC;
 559                 io_ctl_map_page(io_ctl, 0);
 560         }
 561
 562         memcpy(io_ctl->cur, bitmap, PAGE_SIZE);
 563         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 564         if (io_ctl->index < io_ctl->num_pages)
 565                 io_ctl_map_page(io_ctl, 0);
 566         return 0;
 567 }
 568
 569 static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
 570 {
 571         /*
 572          * If we're not on the boundary we know we've modified the page and we
 573          * need to crc the page.
 574          */
 575         if (io_ctl->cur != io_ctl->orig)
 576                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 577         else
 578                 io_ctl_unmap_page(io_ctl);
 579
 580         while (io_ctl->index < io_ctl->num_pages) {
 581                 io_ctl_map_page(io_ctl, 1);
 582                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 583         }
 584 }
 585
 586 static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
 587                             struct btrfs_free_space *entry, u8 *type)
 588 {
 589         struct btrfs_free_space_entry *e;
 590         int ret;
 591
 592         if (!io_ctl->cur) {
 593                 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 594                 if (ret)
 595                         return ret;
 596         }
 597
 598         e = io_ctl->cur;
 599         entry->offset = le64_to_cpu(e->offset);
 600         entry->bytes = le64_to_cpu(e->bytes);
 601         *type = e->type;
 602         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 603         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 604
 605         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 606                 return 0;
 607
 608         io_ctl_unmap_page(io_ctl);
 609
 610         return 0;
 611 }
 612
 613 static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
 614                               struct btrfs_free_space *entry)
 615 {
 616         int ret;
 617
 618         ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 619         if (ret)
 620                 return ret;
 621
 622         memcpy(entry->bitmap, io_ctl->cur, PAGE_SIZE);
 623         io_ctl_unmap_page(io_ctl);
 624
 625         return 0;
 626 }
 627
 628 /*
 629  * Since we attach pinned extents after the fact we can have contiguous sections
 630  * of free space that are split up in entries.  This poses a problem with the
 631  * tree logging stuff since it could have allocated across what appears to be 2
 632  * entries since we would have merged the entries when adding the pinned extents
 633  * back to the free space cache.  So run through the space cache that we just
 634  * loaded and merge contiguous entries.  This will make the log replay stuff not
 635  * blow up and it will make for nicer allocator behavior.
 636  */
 637 static void merge_space_tree(struct btrfs_free_space_ctl *ctl)
 638 {
 639         struct btrfs_free_space *e, *prev = NULL;
 640         struct rb_node *n;
 641
 642 again:
 643         spin_lock(&ctl->tree_lock);
 644         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
 645                 e = rb_entry(n, struct btrfs_free_space, offset_index);
 646                 if (!prev)
 647                         goto next;
 648                 if (e->bitmap || prev->bitmap)
 649                         goto next;
 650                 if (prev->offset + prev->bytes == e->offset) {
 651                         unlink_free_space(ctl, prev);
 652                         unlink_free_space(ctl, e);
 653                         prev->bytes += e->bytes;
 654                         kmem_cache_free(btrfs_free_space_cachep, e);
 655                         link_free_space(ctl, prev);
 656                         prev = NULL;
 657                         spin_unlock(&ctl->tree_lock);
 658                         goto again;
 659                 }
 660 next:
 661                 prev = e;
 662         }
 663         spin_unlock(&ctl->tree_lock);
 664 }
 665
 666 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
 667                                    struct btrfs_free_space_ctl *ctl,
 668                                    struct btrfs_path *path, u64 offset)
 669 {
 670         struct btrfs_free_space_header *header;
 671         struct extent_buffer *leaf;
 672         struct btrfs_io_ctl io_ctl;
 673         struct btrfs_key key;
 674         struct btrfs_free_space *e, *n;
 675         LIST_HEAD(bitmaps);
 676         u64 num_entries;
 677         u64 num_bitmaps;
 678         u64 generation;
 679         u8 type;
 680         int ret = 0;
 681
 682         /* Nothing in the space cache, goodbye */
 683         if (!i_size_read(inode))
 684                 return 0;
 685
 686         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 687         key.offset = offset;
 688         key.type = 0;
 689
 690         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 691         if (ret < 0)
 692                 return 0;
 693         else if (ret > 0) {
 694                 btrfs_release_path(path);
 695                 return 0;
 696         }
 697
 698         ret = -1;
 699
 700         leaf = path->nodes[0];
 701         header = btrfs_item_ptr(leaf, path->slots[0],
 702                                 struct btrfs_free_space_header);
 703         num_entries = btrfs_free_space_entries(leaf, header);
 704         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 705         generation = btrfs_free_space_generation(leaf, header);
 706         btrfs_release_path(path);
 707
 708         if (!BTRFS_I(inode)->generation) {
 709                 btrfs_info(root->fs_info,
 710                            "The free space cache file (%llu) is invalid. skip it\n",
 711                            offset);
 712                 return 0;
 713         }
 714
 715         if (BTRFS_I(inode)->generation != generation) {
 716                 btrfs_err(root->fs_info,
 717                         "free space inode generation (%llu) did not match free space cache generation (%llu)",
 718                         BTRFS_I(inode)->generation, generation);
 719                 return 0;
 720         }
 721
 722         if (!num_entries)
 723                 return 0;
 724
 725         ret = io_ctl_init(&io_ctl, inode, root, 0);
 726         if (ret)
 727                 return ret;
 728
 729         ret = readahead_cache(inode);
 730         if (ret)
 731                 goto out;
 732
 733         ret = io_ctl_prepare_pages(&io_ctl, inode, 1);
 734         if (ret)
 735                 goto out;
 736
 737         ret = io_ctl_check_crc(&io_ctl, 0);
 738         if (ret)
 739                 goto free_cache;
 740
 741         ret = io_ctl_check_generation(&io_ctl, generation);
 742         if (ret)
 743                 goto free_cache;
 744
 745         while (num_entries) {
 746                 e = kmem_cache_zalloc(btrfs_free_space_cachep,
 747                                       GFP_NOFS);
 748                 if (!e)
 749                         goto free_cache;
 750
 751                 ret = io_ctl_read_entry(&io_ctl, e, &type);
 752                 if (ret) {
 753                         kmem_cache_free(btrfs_free_space_cachep, e);
 754                         goto free_cache;
 755                 }
 756
 757                 if (!e->bytes) {
 758                         kmem_cache_free(btrfs_free_space_cachep, e);
 759                         goto free_cache;
 760                 }
 761
 762                 if (type == BTRFS_FREE_SPACE_EXTENT) {
 763                         spin_lock(&ctl->tree_lock);
 764                         ret = link_free_space(ctl, e);
 765                         spin_unlock(&ctl->tree_lock);
 766                         if (ret) {
 767                                 btrfs_err(root->fs_info,
 768                                         "Duplicate entries in free space cache, dumping");
 769                                 kmem_cache_free(btrfs_free_space_cachep, e);
 770                                 goto free_cache;
 771                         }
 772                 } else {
 773                         ASSERT(num_bitmaps);
 774                         num_bitmaps--;
 775                         e->bitmap = kzalloc(PAGE_SIZE, GFP_NOFS);
 776                         if (!e->bitmap) {
 777                                 kmem_cache_free(
 778                                         btrfs_free_space_cachep, e);
 779                                 goto free_cache;
 780                         }
 781                         spin_lock(&ctl->tree_lock);
 782                         ret = link_free_space(ctl, e);
 783                         ctl->total_bitmaps++;
 784                         ctl->op->recalc_thresholds(ctl);
 785                         spin_unlock(&ctl->tree_lock);
 786                         if (ret) {
 787                                 btrfs_err(root->fs_info,
 788                                         "Duplicate entries in free space cache, dumping");
 789                                 kmem_cache_free(btrfs_free_space_cachep, e);
 790                                 goto free_cache;
 791                         }
 792                         list_add_tail(&e->list, &bitmaps);
 793                 }
 794
 795                 num_entries--;
 796         }
 797
 798         io_ctl_unmap_page(&io_ctl);
 799
 800         /*
 801          * We add the bitmaps at the end of the entries in order that
 802          * the bitmap entries are added to the cache.
 803          */
 804         list_for_each_entry_safe(e, n, &bitmaps, list) {
 805                 list_del_init(&e->list);
 806                 ret = io_ctl_read_bitmap(&io_ctl, e);
 807                 if (ret)
 808                         goto free_cache;
 809         }
 810
 811         io_ctl_drop_pages(&io_ctl);
 812         merge_space_tree(ctl);
 813         ret = 1;
 814 out:
 815         io_ctl_free(&io_ctl);
 816         return ret;
 817 free_cache:
 818         io_ctl_drop_pages(&io_ctl);
 819         __btrfs_remove_free_space_cache(ctl);
 820         goto out;
 821 }
 822
 823 int load_free_space_cache(struct btrfs_fs_info *fs_info,
 824                           struct btrfs_block_group_cache *block_group)
 825 {
 826         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
 827         struct btrfs_root *root = fs_info->tree_root;
 828         struct inode *inode;
 829         struct btrfs_path *path;
 830         int ret = 0;
 831         bool matched;
 832         u64 used = btrfs_block_group_used(&block_group->item);
 833
 834         /*
 835          * If this block group has been marked to be cleared for one reason or
 836          * another then we can't trust the on disk cache, so just return.
 837          */
 838         spin_lock(&block_group->lock);
 839         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 840                 spin_unlock(&block_group->lock);
 841                 return 0;
 842         }
 843         spin_unlock(&block_group->lock);
 844
 845         path = btrfs_alloc_path();
 846         if (!path)
 847                 return 0;
 848         path->search_commit_root = 1;
 849         path->skip_locking = 1;
 850
 851         inode = lookup_free_space_inode(root, block_group, path);
 852         if (IS_ERR(inode)) {
 853                 btrfs_free_path(path);
 854                 return 0;
 855         }
 856
 857         /* We may have converted the inode and made the cache invalid. */
 858         spin_lock(&block_group->lock);
 859         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 860                 spin_unlock(&block_group->lock);
 861                 btrfs_free_path(path);
 862                 goto out;
 863         }
 864         spin_unlock(&block_group->lock);
 865
 866         ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
 867                                       path, block_group->key.objectid);
 868         btrfs_free_path(path);
 869         if (ret <= 0)
 870                 goto out;
 871
 872         spin_lock(&ctl->tree_lock);
 873         matched = (ctl->free_space == (block_group->key.offset - used -
 874                                        block_group->bytes_super));
 875         spin_unlock(&ctl->tree_lock);
 876
 877         if (!matched) {
 878                 __btrfs_remove_free_space_cache(ctl);
 879                 btrfs_warn(fs_info,
 880                            "block group %llu has wrong amount of free space",
 881                            block_group->key.objectid);
 882                 ret = -1;
 883         }
 884 out:
 885         if (ret < 0) {
 886                 /* This cache is bogus, make sure it gets cleared */
 887                 spin_lock(&block_group->lock);
 888                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 889                 spin_unlock(&block_group->lock);
 890                 ret = 0;
 891
 892                 btrfs_warn(fs_info,
 893                            "failed to load free space cache for block group %llu, rebuilding it now",
 894                            block_group->key.objectid);
 895         }
 896
 897         iput(inode);
 898         return ret;
 899 }
 900
 901 static noinline_for_stack
 902 int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
 903                               struct btrfs_free_space_ctl *ctl,
 904                               struct btrfs_block_group_cache *block_group,
 905                               int *entries, int *bitmaps,
 906                               struct list_head *bitmap_list)
 907 {
 908         int ret;
 909         struct btrfs_free_cluster *cluster = NULL;
 910         struct btrfs_free_cluster *cluster_locked = NULL;
 911         struct rb_node *node = rb_first(&ctl->free_space_offset);
 912         struct btrfs_trim_range *trim_entry;
 913
 914         /* Get the cluster for this block_group if it exists */
 915         if (block_group && !list_empty(&block_group->cluster_list)) {
 916                 cluster = list_entry(block_group->cluster_list.next,
 917                                      struct btrfs_free_cluster,
 918                                      block_group_list);
 919         }
 920
 921         if (!node && cluster) {
 922                 cluster_locked = cluster;
 923                 spin_lock(&cluster_locked->lock);
 924                 node = rb_first(&cluster->root);
 925                 cluster = NULL;
 926         }
 927
 928         /* Write out the extent entries */
 929         while (node) {
 930                 struct btrfs_free_space *e;
 931
 932                 e = rb_entry(node, struct btrfs_free_space, offset_index);
 933                 *entries += 1;
 934
 935                 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
 936                                        e->bitmap);
 937                 if (ret)
 938                         goto fail;
 939
 940                 if (e->bitmap) {
 941                         list_add_tail(&e->list, bitmap_list);
 942                         *bitmaps += 1;
 943                 }
 944                 node = rb_next(node);
 945                 if (!node && cluster) {
 946                         node = rb_first(&cluster->root);
 947                         cluster_locked = cluster;
 948                         spin_lock(&cluster_locked->lock);
 949                         cluster = NULL;
 950                 }
 951         }
 952         if (cluster_locked) {
 953                 spin_unlock(&cluster_locked->lock);
 954                 cluster_locked = NULL;
 955         }
 956
 957         /*
 958          * Make sure we don't miss any range that was removed from our rbtree
 959          * because trimming is running. Otherwise after a umount+mount (or crash
 960          * after committing the transaction) we would leak free space and get
 961          * an inconsistent free space cache report from fsck.
 962          */
 963         list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
 964                 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
 965                                        trim_entry->bytes, NULL);
 966                 if (ret)
 967                         goto fail;
 968                 *entries += 1;
 969         }
 970
 971         return 0;
 972 fail:
 973         if (cluster_locked)
 974                 spin_unlock(&cluster_locked->lock);
 975         return -ENOSPC;
 976 }
 977
 978 static noinline_for_stack int
 979 update_cache_item(struct btrfs_trans_handle *trans,
 980                   struct btrfs_root *root,
 981                   struct inode *inode,
 982                   struct btrfs_path *path, u64 offset,
 983                   int entries, int bitmaps)
 984 {
 985         struct btrfs_key key;
 986         struct btrfs_free_space_header *header;
 987         struct extent_buffer *leaf;
 988         int ret;
 989
 990         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 991         key.offset = offset;
 992         key.type = 0;
 993
 994         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
 995         if (ret < 0) {
 996                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
 997                                  EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL,
 998                                  GFP_NOFS);
 999                 goto fail;
1000         }
1001         leaf = path->nodes[0];
1002         if (ret > 0) {
1003                 struct btrfs_key found_key;
1004                 ASSERT(path->slots[0]);
1005                 path->slots[0]--;
1006                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1007                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1008                     found_key.offset != offset) {
1009                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1010                                          inode->i_size - 1,
1011                                          EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0,
1012                                          NULL, GFP_NOFS);
1013                         btrfs_release_path(path);
1014                         goto fail;
1015                 }
1016         }
1017
1018         BTRFS_I(inode)->generation = trans->transid;
1019         header = btrfs_item_ptr(leaf, path->slots[0],
1020                                 struct btrfs_free_space_header);
1021         btrfs_set_free_space_entries(leaf, header, entries);
1022         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1023         btrfs_set_free_space_generation(leaf, header, trans->transid);
1024         btrfs_mark_buffer_dirty(leaf);
1025         btrfs_release_path(path);
1026
1027         return 0;
1028
1029 fail:
1030         return -1;
1031 }
1032
1033 static noinline_for_stack int
1034 write_pinned_extent_entries(struct btrfs_root *root,
1035                             struct btrfs_block_group_cache *block_group,
1036                             struct btrfs_io_ctl *io_ctl,
1037                             int *entries)
1038 {
1039         u64 start, extent_start, extent_end, len;
1040         struct extent_io_tree *unpin = NULL;
1041         int ret;
1042
1043         if (!block_group)
1044                 return 0;
1045
1046         /*
1047          * We want to add any pinned extents to our free space cache
1048          * so we don't leak the space
1049          *
1050          * We shouldn't have switched the pinned extents yet so this is the
1051          * right one
1052          */
1053         unpin = root->fs_info->pinned_extents;
1054
1055         start = block_group->key.objectid;
1056
1057         while (start < block_group->key.objectid + block_group->key.offset) {
1058                 ret = find_first_extent_bit(unpin, start,
1059                                             &extent_start, &extent_end,
1060                                             EXTENT_DIRTY, NULL);
1061                 if (ret)
1062                         return 0;
1063
1064                 /* This pinned extent is out of our range */
1065                 if (extent_start >= block_group->key.objectid +
1066                     block_group->key.offset)
1067                         return 0;
1068
1069                 extent_start = max(extent_start, start);
1070                 extent_end = min(block_group->key.objectid +
1071                                  block_group->key.offset, extent_end + 1);
1072                 len = extent_end - extent_start;
1073
1074                 *entries += 1;
1075                 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1076                 if (ret)
1077                         return -ENOSPC;
1078
1079                 start = extent_end;
1080         }
1081
1082         return 0;
1083 }
1084
1085 static noinline_for_stack int
1086 write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1087 {
1088         struct btrfs_free_space *entry, *next;
1089         int ret;
1090
1091         /* Write out the bitmaps */
1092         list_for_each_entry_safe(entry, next, bitmap_list, list) {
1093                 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1094                 if (ret)
1095                         return -ENOSPC;
1096                 list_del_init(&entry->list);
1097         }
1098
1099         return 0;
1100 }
1101
1102 static int flush_dirty_cache(struct inode *inode)
1103 {
1104         int ret;
1105
1106         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1107         if (ret)
1108                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1109                                  EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL,
1110                                  GFP_NOFS);
1111
1112         return ret;
1113 }
1114
1115 static void noinline_for_stack
1116 cleanup_bitmap_list(struct list_head *bitmap_list)
1117 {
1118         struct btrfs_free_space *entry, *next;
1119
1120         list_for_each_entry_safe(entry, next, bitmap_list, list)
1121                 list_del_init(&entry->list);
1122 }
1123
1124 static void noinline_for_stack
1125 cleanup_write_cache_enospc(struct inode *inode,
1126                            struct btrfs_io_ctl *io_ctl,
1127                            struct extent_state **cached_state,
1128                            struct list_head *bitmap_list)
1129 {
1130         io_ctl_drop_pages(io_ctl);
1131         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1132                              i_size_read(inode) - 1, cached_state,
1133                              GFP_NOFS);
1134 }
1135
1136 int btrfs_wait_cache_io(struct btrfs_root *root,
1137                         struct btrfs_trans_handle *trans,
1138                         struct btrfs_block_group_cache *block_group,
1139                         struct btrfs_io_ctl *io_ctl,
1140                         struct btrfs_path *path, u64 offset)
1141 {
1142         int ret;
1143         struct inode *inode = io_ctl->inode;
1144
1145         if (!inode)
1146                 return 0;
1147
1148         if (block_group)
1149                 root = root->fs_info->tree_root;
1150
1151         /* Flush the dirty pages in the cache file. */
1152         ret = flush_dirty_cache(inode);
1153         if (ret)
1154                 goto out;
1155
1156         /* Update the cache item to tell everyone this cache file is valid. */
1157         ret = update_cache_item(trans, root, inode, path, offset,
1158                                 io_ctl->entries, io_ctl->bitmaps);
1159 out:
1160         io_ctl_free(io_ctl);
1161         if (ret) {
1162                 invalidate_inode_pages2(inode->i_mapping);
1163                 BTRFS_I(inode)->generation = 0;
1164                 if (block_group) {
1165 #ifdef DEBUG
1166                         btrfs_err(root->fs_info,
1167                                 "failed to write free space cache for block group %llu",
1168                                 block_group->key.objectid);
1169 #endif
1170                 }
1171         }
1172         btrfs_update_inode(trans, root, inode);
1173
1174         if (block_group) {
1175                 /* the dirty list is protected by the dirty_bgs_lock */
1176                 spin_lock(&trans->transaction->dirty_bgs_lock);
1177
1178                 /* the disk_cache_state is protected by the block group lock */
1179                 spin_lock(&block_group->lock);
1180
1181                 /*
1182                  * only mark this as written if we didn't get put back on
1183                  * the dirty list while waiting for IO.   Otherwise our
1184                  * cache state won't be right, and we won't get written again
1185                  */
1186                 if (!ret && list_empty(&block_group->dirty_list))
1187                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1188                 else if (ret)
1189                         block_group->disk_cache_state = BTRFS_DC_ERROR;
1190
1191                 spin_unlock(&block_group->lock);
1192                 spin_unlock(&trans->transaction->dirty_bgs_lock);
1193                 io_ctl->inode = NULL;
1194                 iput(inode);
1195         }
1196
1197         return ret;
1198
1199 }
1200
1201 /**
1202  * __btrfs_write_out_cache - write out cached info to an inode
1203  * @root - the root the inode belongs to
1204  * @ctl - the free space cache we are going to write out
1205  * @block_group - the block_group for this cache if it belongs to a block_group
1206  * @trans - the trans handle
1207  * @path - the path to use
1208  * @offset - the offset for the key we'll insert
1209  *
1210  * This function writes out a free space cache struct to disk for quick recovery
1211  * on mount.  This will return 0 if it was successful in writing the cache out,
1212  * or an errno if it was not.
1213  */
1214 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1215                                    struct btrfs_free_space_ctl *ctl,
1216                                    struct btrfs_block_group_cache *block_group,
1217                                    struct btrfs_io_ctl *io_ctl,
1218                                    struct btrfs_trans_handle *trans,
1219                                    struct btrfs_path *path, u64 offset)
1220 {
1221         struct extent_state *cached_state = NULL;
1222         LIST_HEAD(bitmap_list);
1223         int entries = 0;
1224         int bitmaps = 0;
1225         int ret;
1226         int must_iput = 0;
1227
1228         if (!i_size_read(inode))
1229                 return -EIO;
1230
1231         WARN_ON(io_ctl->pages);
1232         ret = io_ctl_init(io_ctl, inode, root, 1);
1233         if (ret)
1234                 return ret;
1235
1236         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1237                 down_write(&block_group->data_rwsem);
1238                 spin_lock(&block_group->lock);
1239                 if (block_group->delalloc_bytes) {
1240                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1241                         spin_unlock(&block_group->lock);
1242                         up_write(&block_group->data_rwsem);
1243                         BTRFS_I(inode)->generation = 0;
1244                         ret = 0;
1245                         must_iput = 1;
1246                         goto out;
1247                 }
1248                 spin_unlock(&block_group->lock);
1249         }
1250
1251         /* Lock all pages first so we can lock the extent safely. */
1252         ret = io_ctl_prepare_pages(io_ctl, inode, 0);
1253         if (ret)
1254                 goto out;
1255
1256         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1257                          &cached_state);
1258
1259         io_ctl_set_generation(io_ctl, trans->transid);
1260
1261         mutex_lock(&ctl->cache_writeout_mutex);
1262         /* Write out the extent entries in the free space cache */
1263         spin_lock(&ctl->tree_lock);
1264         ret = write_cache_extent_entries(io_ctl, ctl,
1265                                          block_group, &entries, &bitmaps,
1266                                          &bitmap_list);
1267         if (ret)
1268                 goto out_nospc_locked;
1269
1270         /*
1271          * Some spaces that are freed in the current transaction are pinned,
1272          * they will be added into free space cache after the transaction is
1273          * committed, we shouldn't lose them.
1274          *
1275          * If this changes while we are working we'll get added back to
1276          * the dirty list and redo it.  No locking needed
1277          */
1278         ret = write_pinned_extent_entries(root, block_group, io_ctl, &entries);
1279         if (ret)
1280                 goto out_nospc_locked;
1281
1282         /*
1283          * At last, we write out all the bitmaps and keep cache_writeout_mutex
1284          * locked while doing it because a concurrent trim can be manipulating
1285          * or freeing the bitmap.
1286          */
1287         ret = write_bitmap_entries(io_ctl, &bitmap_list);
1288         spin_unlock(&ctl->tree_lock);
1289         mutex_unlock(&ctl->cache_writeout_mutex);
1290         if (ret)
1291                 goto out_nospc;
1292
1293         /* Zero out the rest of the pages just to make sure */
1294         io_ctl_zero_remaining_pages(io_ctl);
1295
1296         /* Everything is written out, now we dirty the pages in the file. */
1297         ret = btrfs_dirty_pages(root, inode, io_ctl->pages, io_ctl->num_pages,
1298                                 0, i_size_read(inode), &cached_state);
1299         if (ret)
1300                 goto out_nospc;
1301
1302         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1303                 up_write(&block_group->data_rwsem);
1304         /*
1305          * Release the pages and unlock the extent, we will flush
1306          * them out later
1307          */
1308         io_ctl_drop_pages(io_ctl);
1309
1310         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1311                              i_size_read(inode) - 1, &cached_state, GFP_NOFS);
1312
1313         /*
1314          * at this point the pages are under IO and we're happy,
1315          * The caller is responsible for waiting on them and updating the
1316          * the cache and the inode
1317          */
1318         io_ctl->entries = entries;
1319         io_ctl->bitmaps = bitmaps;
1320
1321         ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1322         if (ret)
1323                 goto out;
1324
1325         return 0;
1326
1327 out:
1328         io_ctl->inode = NULL;
1329         io_ctl_free(io_ctl);
1330         if (ret) {
1331                 invalidate_inode_pages2(inode->i_mapping);
1332                 BTRFS_I(inode)->generation = 0;
1333         }
1334         btrfs_update_inode(trans, root, inode);
1335         if (must_iput)
1336                 iput(inode);
1337         return ret;
1338
1339 out_nospc_locked:
1340         cleanup_bitmap_list(&bitmap_list);
1341         spin_unlock(&ctl->tree_lock);
1342         mutex_unlock(&ctl->cache_writeout_mutex);
1343
1344 out_nospc:
1345         cleanup_write_cache_enospc(inode, io_ctl, &cached_state, &bitmap_list);
1346
1347         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1348                 up_write(&block_group->data_rwsem);
1349
1350         goto out;
1351 }
1352
1353 int btrfs_write_out_cache(struct btrfs_fs_info *fs_info,
1354                           struct btrfs_trans_handle *trans,
1355                           struct btrfs_block_group_cache *block_group,
1356                           struct btrfs_path *path)
1357 {
1358         struct btrfs_root *root = fs_info->tree_root;
1359         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1360         struct inode *inode;
1361         int ret = 0;
1362
1363         spin_lock(&block_group->lock);
1364         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1365                 spin_unlock(&block_group->lock);
1366                 return 0;
1367         }
1368         spin_unlock(&block_group->lock);
1369
1370         inode = lookup_free_space_inode(root, block_group, path);
1371         if (IS_ERR(inode))
1372                 return 0;
1373
1374         ret = __btrfs_write_out_cache(root, inode, ctl, block_group,
1375                                       &block_group->io_ctl, trans,
1376                                       path, block_group->key.objectid);
1377         if (ret) {
1378 #ifdef DEBUG
1379                 btrfs_err(root->fs_info,
1380                         "failed to write free space cache for block group %llu",
1381                         block_group->key.objectid);
1382 #endif
1383                 spin_lock(&block_group->lock);
1384                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1385                 spin_unlock(&block_group->lock);
1386
1387                 block_group->io_ctl.inode = NULL;
1388                 iput(inode);
1389         }
1390
1391         /*
1392          * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1393          * to wait for IO and put the inode
1394          */
1395
1396         return ret;
1397 }
1398
1399 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1400                                           u64 offset)
1401 {
1402         ASSERT(offset >= bitmap_start);
1403         offset -= bitmap_start;
1404         return (unsigned long)(div_u64(offset, unit));
1405 }
1406
1407 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1408 {
1409         return (unsigned long)(div_u64(bytes, unit));
1410 }
1411
1412 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1413                                    u64 offset)
1414 {
1415         u64 bitmap_start;
1416         u64 bytes_per_bitmap;
1417
1418         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1419         bitmap_start = offset - ctl->start;
1420         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1421         bitmap_start *= bytes_per_bitmap;
1422         bitmap_start += ctl->start;
1423
1424         return bitmap_start;
1425 }
1426
1427 static int tree_insert_offset(struct rb_root *root, u64 offset,
1428                               struct rb_node *node, int bitmap)
1429 {
1430         struct rb_node **p = &root->rb_node;
1431         struct rb_node *parent = NULL;
1432         struct btrfs_free_space *info;
1433
1434         while (*p) {
1435                 parent = *p;
1436                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1437
1438                 if (offset < info->offset) {
1439                         p = &(*p)->rb_left;
1440                 } else if (offset > info->offset) {
1441                         p = &(*p)->rb_right;
1442                 } else {
1443                         /*
1444                          * we could have a bitmap entry and an extent entry
1445                          * share the same offset.  If this is the case, we want
1446                          * the extent entry to always be found first if we do a
1447                          * linear search through the tree, since we want to have
1448                          * the quickest allocation time, and allocating from an
1449                          * extent is faster than allocating from a bitmap.  So
1450                          * if we're inserting a bitmap and we find an entry at
1451                          * this offset, we want to go right, or after this entry
1452                          * logically.  If we are inserting an extent and we've
1453                          * found a bitmap, we want to go left, or before
1454                          * logically.
1455                          */
1456                         if (bitmap) {
1457                                 if (info->bitmap) {
1458                                         WARN_ON_ONCE(1);
1459                                         return -EEXIST;
1460                                 }
1461                                 p = &(*p)->rb_right;
1462                         } else {
1463                                 if (!info->bitmap) {
1464                                         WARN_ON_ONCE(1);
1465                                         return -EEXIST;
1466                                 }
1467                                 p = &(*p)->rb_left;
1468                         }
1469                 }
1470         }
1471
1472         rb_link_node(node, parent, p);
1473         rb_insert_color(node, root);
1474
1475         return 0;
1476 }
1477
1478 /*
1479  * searches the tree for the given offset.
1480  *
1481  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1482  * want a section that has at least bytes size and comes at or after the given
1483  * offset.
1484  */
1485 static struct btrfs_free_space *
1486 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1487                    u64 offset, int bitmap_only, int fuzzy)
1488 {
1489         struct rb_node *n = ctl->free_space_offset.rb_node;
1490         struct btrfs_free_space *entry, *prev = NULL;
1491
1492         /* find entry that is closest to the 'offset' */
1493         while (1) {
1494                 if (!n) {
1495                         entry = NULL;
1496                         break;
1497                 }
1498
1499                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1500                 prev = entry;
1501
1502                 if (offset < entry->offset)
1503                         n = n->rb_left;
1504                 else if (offset > entry->offset)
1505                         n = n->rb_right;
1506                 else
1507                         break;
1508         }
1509
1510         if (bitmap_only) {
1511                 if (!entry)
1512                         return NULL;
1513                 if (entry->bitmap)
1514                         return entry;
1515
1516                 /*
1517                  * bitmap entry and extent entry may share same offset,
1518                  * in that case, bitmap entry comes after extent entry.
1519                  */
1520                 n = rb_next(n);
1521                 if (!n)
1522                         return NULL;
1523                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1524                 if (entry->offset != offset)
1525                         return NULL;
1526
1527                 WARN_ON(!entry->bitmap);
1528                 return entry;
1529         } else if (entry) {
1530                 if (entry->bitmap) {
1531                         /*
1532                          * if previous extent entry covers the offset,
1533                          * we should return it instead of the bitmap entry
1534                          */
1535                         n = rb_prev(&entry->offset_index);
1536                         if (n) {
1537                                 prev = rb_entry(n, struct btrfs_free_space,
1538                                                 offset_index);
1539                                 if (!prev->bitmap &&
1540                                     prev->offset + prev->bytes > offset)
1541                                         entry = prev;
1542                         }
1543                 }
1544                 return entry;
1545         }
1546
1547         if (!prev)
1548                 return NULL;
1549
1550         /* find last entry before the 'offset' */
1551         entry = prev;
1552         if (entry->offset > offset) {
1553                 n = rb_prev(&entry->offset_index);
1554                 if (n) {
1555                         entry = rb_entry(n, struct btrfs_free_space,
1556                                         offset_index);
1557                         ASSERT(entry->offset <= offset);
1558                 } else {
1559                         if (fuzzy)
1560                                 return entry;
1561                         else
1562                                 return NULL;
1563                 }
1564         }
1565
1566         if (entry->bitmap) {
1567                 n = rb_prev(&entry->offset_index);
1568                 if (n) {
1569                         prev = rb_entry(n, struct btrfs_free_space,
1570                                         offset_index);
1571                         if (!prev->bitmap &&
1572                             prev->offset + prev->bytes > offset)
1573                                 return prev;
1574                 }
1575                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1576                         return entry;
1577         } else if (entry->offset + entry->bytes > offset)
1578                 return entry;
1579
1580         if (!fuzzy)
1581                 return NULL;
1582
1583         while (1) {
1584                 if (entry->bitmap) {
1585                         if (entry->offset + BITS_PER_BITMAP *
1586                             ctl->unit > offset)
1587                                 break;
1588                 } else {
1589                         if (entry->offset + entry->bytes > offset)
1590                                 break;
1591                 }
1592
1593                 n = rb_next(&entry->offset_index);
1594                 if (!n)
1595                         return NULL;
1596                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1597         }
1598         return entry;
1599 }
1600
1601 static inline void
1602 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1603                     struct btrfs_free_space *info)
1604 {
1605         rb_erase(&info->offset_index, &ctl->free_space_offset);
1606         ctl->free_extents--;
1607 }
1608
1609 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1610                               struct btrfs_free_space *info)
1611 {
1612         __unlink_free_space(ctl, info);
1613         ctl->free_space -= info->bytes;
1614 }
1615
1616 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1617                            struct btrfs_free_space *info)
1618 {
1619         int ret = 0;
1620
1621         ASSERT(info->bytes || info->bitmap);
1622         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1623                                  &info->offset_index, (info->bitmap != NULL));
1624         if (ret)
1625                 return ret;
1626
1627         ctl->free_space += info->bytes;
1628         ctl->free_extents++;
1629         return ret;
1630 }
1631
1632 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1633 {
1634         struct btrfs_block_group_cache *block_group = ctl->private;
1635         u64 max_bytes;
1636         u64 bitmap_bytes;
1637         u64 extent_bytes;
1638         u64 size = block_group->key.offset;
1639         u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
1640         u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
1641
1642         max_bitmaps = max_t(u64, max_bitmaps, 1);
1643
1644         ASSERT(ctl->total_bitmaps <= max_bitmaps);
1645
1646         /*
1647          * The goal is to keep the total amount of memory used per 1gb of space
1648          * at or below 32k, so we need to adjust how much memory we allow to be
1649          * used by extent based free space tracking
1650          */
1651         if (size < SZ_1G)
1652                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1653         else
1654                 max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
1655
1656         /*
1657          * we want to account for 1 more bitmap than what we have so we can make
1658          * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1659          * we add more bitmaps.
1660          */
1661         bitmap_bytes = (ctl->total_bitmaps + 1) * ctl->unit;
1662
1663         if (bitmap_bytes >= max_bytes) {
1664                 ctl->extents_thresh = 0;
1665                 return;
1666         }
1667
1668         /*
1669          * we want the extent entry threshold to always be at most 1/2 the max
1670          * bytes we can have, or whatever is less than that.
1671          */
1672         extent_bytes = max_bytes - bitmap_bytes;
1673         extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
1674
1675         ctl->extents_thresh =
1676                 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
1677 }
1678
1679 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1680                                        struct btrfs_free_space *info,
1681                                        u64 offset, u64 bytes)
1682 {
1683         unsigned long start, count;
1684
1685         start = offset_to_bit(info->offset, ctl->unit, offset);
1686         count = bytes_to_bits(bytes, ctl->unit);
1687         ASSERT(start + count <= BITS_PER_BITMAP);
1688
1689         bitmap_clear(info->bitmap, start, count);
1690
1691         info->bytes -= bytes;
1692 }
1693
1694 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1695                               struct btrfs_free_space *info, u64 offset,
1696                               u64 bytes)
1697 {
1698         __bitmap_clear_bits(ctl, info, offset, bytes);
1699         ctl->free_space -= bytes;
1700 }
1701
1702 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1703                             struct btrfs_free_space *info, u64 offset,
1704                             u64 bytes)
1705 {
1706         unsigned long start, count;
1707
1708         start = offset_to_bit(info->offset, ctl->unit, offset);
1709         count = bytes_to_bits(bytes, ctl->unit);
1710         ASSERT(start + count <= BITS_PER_BITMAP);
1711
1712         bitmap_set(info->bitmap, start, count);
1713
1714         info->bytes += bytes;
1715         ctl->free_space += bytes;
1716 }
1717
1718 /*
1719  * If we can not find suitable extent, we will use bytes to record
1720  * the size of the max extent.
1721  */
1722 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1723                          struct btrfs_free_space *bitmap_info, u64 *offset,
1724                          u64 *bytes, bool for_alloc)
1725 {
1726         unsigned long found_bits = 0;
1727         unsigned long max_bits = 0;
1728         unsigned long bits, i;
1729         unsigned long next_zero;
1730         unsigned long extent_bits;
1731
1732         /*
1733          * Skip searching the bitmap if we don't have a contiguous section that
1734          * is large enough for this allocation.
1735          */
1736         if (for_alloc &&
1737             bitmap_info->max_extent_size &&
1738             bitmap_info->max_extent_size < *bytes) {
1739                 *bytes = bitmap_info->max_extent_size;
1740                 return -1;
1741         }
1742
1743         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1744                           max_t(u64, *offset, bitmap_info->offset));
1745         bits = bytes_to_bits(*bytes, ctl->unit);
1746
1747         for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1748                 if (for_alloc && bits == 1) {
1749                         found_bits = 1;
1750                         break;
1751                 }
1752                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1753                                                BITS_PER_BITMAP, i);
1754                 extent_bits = next_zero - i;
1755                 if (extent_bits >= bits) {
1756                         found_bits = extent_bits;
1757                         break;
1758                 } else if (extent_bits > max_bits) {
1759                         max_bits = extent_bits;
1760                 }
1761                 i = next_zero;
1762         }
1763
1764         if (found_bits) {
1765                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1766                 *bytes = (u64)(found_bits) * ctl->unit;
1767                 return 0;
1768         }
1769
1770         *bytes = (u64)(max_bits) * ctl->unit;
1771         bitmap_info->max_extent_size = *bytes;
1772         return -1;
1773 }
1774
1775 /* Cache the size of the max extent in bytes */
1776 static struct btrfs_free_space *
1777 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1778                 unsigned long align, u64 *max_extent_size)
1779 {
1780         struct btrfs_free_space *entry;
1781         struct rb_node *node;
1782         u64 tmp;
1783         u64 align_off;
1784         int ret;
1785
1786         if (!ctl->free_space_offset.rb_node)
1787                 goto out;
1788
1789         entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1790         if (!entry)
1791                 goto out;
1792
1793         for (node = &entry->offset_index; node; node = rb_next(node)) {
1794                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1795                 if (entry->bytes < *bytes) {
1796                         if (entry->bytes > *max_extent_size)
1797                                 *max_extent_size = entry->bytes;
1798                         continue;
1799                 }
1800
1801                 /* make sure the space returned is big enough
1802                  * to match our requested alignment
1803                  */
1804                 if (*bytes >= align) {
1805                         tmp = entry->offset - ctl->start + align - 1;
1806                         tmp = div64_u64(tmp, align);
1807                         tmp = tmp * align + ctl->start;
1808                         align_off = tmp - entry->offset;
1809                 } else {
1810                         align_off = 0;
1811                         tmp = entry->offset;
1812                 }
1813
1814                 if (entry->bytes < *bytes + align_off) {
1815                         if (entry->bytes > *max_extent_size)
1816                                 *max_extent_size = entry->bytes;
1817                         continue;
1818                 }
1819
1820                 if (entry->bitmap) {
1821                         u64 size = *bytes;
1822
1823                         ret = search_bitmap(ctl, entry, &tmp, &size, true);
1824                         if (!ret) {
1825                                 *offset = tmp;
1826                                 *bytes = size;
1827                                 return entry;
1828                         } else if (size > *max_extent_size) {
1829                                 *max_extent_size = size;
1830                         }
1831                         continue;
1832                 }
1833
1834                 *offset = tmp;
1835                 *bytes = entry->bytes - align_off;
1836                 return entry;
1837         }
1838 out:
1839         return NULL;
1840 }
1841
1842 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1843                            struct btrfs_free_space *info, u64 offset)
1844 {
1845         info->offset = offset_to_bitmap(ctl, offset);
1846         info->bytes = 0;
1847         INIT_LIST_HEAD(&info->list);
1848         link_free_space(ctl, info);
1849         ctl->total_bitmaps++;
1850
1851         ctl->op->recalc_thresholds(ctl);
1852 }
1853
1854 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1855                         struct btrfs_free_space *bitmap_info)
1856 {
1857         unlink_free_space(ctl, bitmap_info);
1858         kfree(bitmap_info->bitmap);
1859         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1860         ctl->total_bitmaps--;
1861         ctl->op->recalc_thresholds(ctl);
1862 }
1863
1864 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1865                               struct btrfs_free_space *bitmap_info,
1866                               u64 *offset, u64 *bytes)
1867 {
1868         u64 end;
1869         u64 search_start, search_bytes;
1870         int ret;
1871
1872 again:
1873         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1874
1875         /*
1876          * We need to search for bits in this bitmap.  We could only cover some
1877          * of the extent in this bitmap thanks to how we add space, so we need
1878          * to search for as much as it as we can and clear that amount, and then
1879          * go searching for the next bit.
1880          */
1881         search_start = *offset;
1882         search_bytes = ctl->unit;
1883         search_bytes = min(search_bytes, end - search_start + 1);
1884         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
1885                             false);
1886         if (ret < 0 || search_start != *offset)
1887                 return -EINVAL;
1888
1889         /* We may have found more bits than what we need */
1890         search_bytes = min(search_bytes, *bytes);
1891
1892         /* Cannot clear past the end of the bitmap */
1893         search_bytes = min(search_bytes, end - search_start + 1);
1894
1895         bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
1896         *offset += search_bytes;
1897         *bytes -= search_bytes;
1898
1899         if (*bytes) {
1900                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1901                 if (!bitmap_info->bytes)
1902                         free_bitmap(ctl, bitmap_info);
1903
1904                 /*
1905                  * no entry after this bitmap, but we still have bytes to
1906                  * remove, so something has gone wrong.
1907                  */
1908                 if (!next)
1909                         return -EINVAL;
1910
1911                 bitmap_info = rb_entry(next, struct btrfs_free_space,
1912                                        offset_index);
1913
1914                 /*
1915                  * if the next entry isn't a bitmap we need to return to let the
1916                  * extent stuff do its work.
1917                  */
1918                 if (!bitmap_info->bitmap)
1919                         return -EAGAIN;
1920
1921                 /*
1922                  * Ok the next item is a bitmap, but it may not actually hold
1923                  * the information for the rest of this free space stuff, so
1924                  * look for it, and if we don't find it return so we can try
1925                  * everything over again.
1926                  */
1927                 search_start = *offset;
1928                 search_bytes = ctl->unit;
1929                 ret = search_bitmap(ctl, bitmap_info, &search_start,
1930                                     &search_bytes, false);
1931                 if (ret < 0 || search_start != *offset)
1932                         return -EAGAIN;
1933
1934                 goto again;
1935         } else if (!bitmap_info->bytes)
1936                 free_bitmap(ctl, bitmap_info);
1937
1938         return 0;
1939 }
1940
1941 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1942                                struct btrfs_free_space *info, u64 offset,
1943                                u64 bytes)
1944 {
1945         u64 bytes_to_set = 0;
1946         u64 end;
1947
1948         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1949
1950         bytes_to_set = min(end - offset, bytes);
1951
1952         bitmap_set_bits(ctl, info, offset, bytes_to_set);
1953
1954         /*
1955          * We set some bytes, we have no idea what the max extent size is
1956          * anymore.
1957          */
1958         info->max_extent_size = 0;
1959
1960         return bytes_to_set;
1961
1962 }
1963
1964 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1965                       struct btrfs_free_space *info)
1966 {
1967         struct btrfs_block_group_cache *block_group = ctl->private;
1968         bool forced = false;
1969
1970 #ifdef CONFIG_BTRFS_DEBUG
1971         if (btrfs_should_fragment_free_space(block_group->fs_info->extent_root,
1972                                              block_group))
1973                 forced = true;
1974 #endif
1975
1976         /*
1977          * If we are below the extents threshold then we can add this as an
1978          * extent, and don't have to deal with the bitmap
1979          */
1980         if (!forced && ctl->free_extents < ctl->extents_thresh) {
1981                 /*
1982                  * If this block group has some small extents we don't want to
1983                  * use up all of our free slots in the cache with them, we want
1984                  * to reserve them to larger extents, however if we have plenty
1985                  * of cache left then go ahead an dadd them, no sense in adding
1986                  * the overhead of a bitmap if we don't have to.
1987                  */
1988                 if (info->bytes <= block_group->sectorsize * 4) {
1989                         if (ctl->free_extents * 2 <= ctl->extents_thresh)
1990                                 return false;
1991                 } else {
1992                         return false;
1993                 }
1994         }
1995
1996         /*
1997          * The original block groups from mkfs can be really small, like 8
1998          * megabytes, so don't bother with a bitmap for those entries.  However
1999          * some block groups can be smaller than what a bitmap would cover but
2000          * are still large enough that they could overflow the 32k memory limit,
2001          * so allow those block groups to still be allowed to have a bitmap
2002          * entry.
2003          */
2004         if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->key.offset)
2005                 return false;
2006
2007         return true;
2008 }
2009
2010 static const struct btrfs_free_space_op free_space_op = {
2011         .recalc_thresholds      = recalculate_thresholds,
2012         .use_bitmap             = use_bitmap,
2013 };
2014
2015 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2016                               struct btrfs_free_space *info)
2017 {
2018         struct btrfs_free_space *bitmap_info;
2019         struct btrfs_block_group_cache *block_group = NULL;
2020         int added = 0;
2021         u64 bytes, offset, bytes_added;
2022         int ret;
2023
2024         bytes = info->bytes;
2025         offset = info->offset;
2026
2027         if (!ctl->op->use_bitmap(ctl, info))
2028                 return 0;
2029
2030         if (ctl->op == &free_space_op)
2031                 block_group = ctl->private;
2032 again:
2033         /*
2034          * Since we link bitmaps right into the cluster we need to see if we
2035          * have a cluster here, and if so and it has our bitmap we need to add
2036          * the free space to that bitmap.
2037          */
2038         if (block_group && !list_empty(&block_group->cluster_list)) {
2039                 struct btrfs_free_cluster *cluster;
2040                 struct rb_node *node;
2041                 struct btrfs_free_space *entry;
2042
2043                 cluster = list_entry(block_group->cluster_list.next,
2044                                      struct btrfs_free_cluster,
2045                                      block_group_list);
2046                 spin_lock(&cluster->lock);
2047                 node = rb_first(&cluster->root);
2048                 if (!node) {
2049                         spin_unlock(&cluster->lock);
2050                         goto no_cluster_bitmap;
2051                 }
2052
2053                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2054                 if (!entry->bitmap) {
2055                         spin_unlock(&cluster->lock);
2056                         goto no_cluster_bitmap;
2057                 }
2058
2059                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2060                         bytes_added = add_bytes_to_bitmap(ctl, entry,
2061                                                           offset, bytes);
2062                         bytes -= bytes_added;
2063                         offset += bytes_added;
2064                 }
2065                 spin_unlock(&cluster->lock);
2066                 if (!bytes) {
2067                         ret = 1;
2068                         goto out;
2069                 }
2070         }
2071
2072 no_cluster_bitmap:
2073         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2074                                          1, 0);
2075         if (!bitmap_info) {
2076                 ASSERT(added == 0);
2077                 goto new_bitmap;
2078         }
2079
2080         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
2081         bytes -= bytes_added;
2082         offset += bytes_added;
2083         added = 0;
2084
2085         if (!bytes) {
2086                 ret = 1;
2087                 goto out;
2088         } else
2089                 goto again;
2090
2091 new_bitmap:
2092         if (info && info->bitmap) {
2093                 add_new_bitmap(ctl, info, offset);
2094                 added = 1;
2095                 info = NULL;
2096                 goto again;
2097         } else {
2098                 spin_unlock(&ctl->tree_lock);
2099
2100                 /* no pre-allocated info, allocate a new one */
2101                 if (!info) {
2102                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
2103                                                  GFP_NOFS);
2104                         if (!info) {
2105                                 spin_lock(&ctl->tree_lock);
2106                                 ret = -ENOMEM;
2107                                 goto out;
2108                         }
2109                 }
2110
2111                 /* allocate the bitmap */
2112                 info->bitmap = kzalloc(PAGE_SIZE, GFP_NOFS);
2113                 spin_lock(&ctl->tree_lock);
2114                 if (!info->bitmap) {
2115                         ret = -ENOMEM;
2116                         goto out;
2117                 }
2118                 goto again;
2119         }
2120
2121 out:
2122         if (info) {
2123                 if (info->bitmap)
2124                         kfree(info->bitmap);
2125                 kmem_cache_free(btrfs_free_space_cachep, info);
2126         }
2127
2128         return ret;
2129 }
2130
2131 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2132                           struct btrfs_free_space *info, bool update_stat)
2133 {
2134         struct btrfs_free_space *left_info;
2135         struct btrfs_free_space *right_info;
2136         bool merged = false;
2137         u64 offset = info->offset;
2138         u64 bytes = info->bytes;
2139
2140         /*
2141          * first we want to see if there is free space adjacent to the range we
2142          * are adding, if there is remove that struct and add a new one to
2143          * cover the entire range
2144          */
2145         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2146         if (right_info && rb_prev(&right_info->offset_index))
2147                 left_info = rb_entry(rb_prev(&right_info->offset_index),
2148                                      struct btrfs_free_space, offset_index);
2149         else
2150                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2151
2152         if (right_info && !right_info->bitmap) {
2153                 if (update_stat)
2154                         unlink_free_space(ctl, right_info);
2155                 else
2156                         __unlink_free_space(ctl, right_info);
2157                 info->bytes += right_info->bytes;
2158                 kmem_cache_free(btrfs_free_space_cachep, right_info);
2159                 merged = true;
2160         }
2161
2162         if (left_info && !left_info->bitmap &&
2163             left_info->offset + left_info->bytes == offset) {
2164                 if (update_stat)
2165                         unlink_free_space(ctl, left_info);
2166                 else
2167                         __unlink_free_space(ctl, left_info);
2168                 info->offset = left_info->offset;
2169                 info->bytes += left_info->bytes;
2170                 kmem_cache_free(btrfs_free_space_cachep, left_info);
2171                 merged = true;
2172         }
2173
2174         return merged;
2175 }
2176
2177 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2178                                      struct btrfs_free_space *info,
2179                                      bool update_stat)
2180 {
2181         struct btrfs_free_space *bitmap;
2182         unsigned long i;
2183         unsigned long j;
2184         const u64 end = info->offset + info->bytes;
2185         const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2186         u64 bytes;
2187
2188         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2189         if (!bitmap)
2190                 return false;
2191
2192         i = offset_to_bit(bitmap->offset, ctl->unit, end);
2193         j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2194         if (j == i)
2195                 return false;
2196         bytes = (j - i) * ctl->unit;
2197         info->bytes += bytes;
2198
2199         if (update_stat)
2200                 bitmap_clear_bits(ctl, bitmap, end, bytes);
2201         else
2202                 __bitmap_clear_bits(ctl, bitmap, end, bytes);
2203
2204         if (!bitmap->bytes)
2205                 free_bitmap(ctl, bitmap);
2206
2207         return true;
2208 }
2209
2210 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2211                                        struct btrfs_free_space *info,
2212                                        bool update_stat)
2213 {
2214         struct btrfs_free_space *bitmap;
2215         u64 bitmap_offset;
2216         unsigned long i;
2217         unsigned long j;
2218         unsigned long prev_j;
2219         u64 bytes;
2220
2221         bitmap_offset = offset_to_bitmap(ctl, info->offset);
2222         /* If we're on a boundary, try the previous logical bitmap. */
2223         if (bitmap_offset == info->offset) {
2224                 if (info->offset == 0)
2225                         return false;
2226                 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2227         }
2228
2229         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2230         if (!bitmap)
2231                 return false;
2232
2233         i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2234         j = 0;
2235         prev_j = (unsigned long)-1;
2236         for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2237                 if (j > i)
2238                         break;
2239                 prev_j = j;
2240         }
2241         if (prev_j == i)
2242                 return false;
2243
2244         if (prev_j == (unsigned long)-1)
2245                 bytes = (i + 1) * ctl->unit;
2246         else
2247                 bytes = (i - prev_j) * ctl->unit;
2248
2249         info->offset -= bytes;
2250         info->bytes += bytes;
2251
2252         if (update_stat)
2253                 bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2254         else
2255                 __bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2256
2257         if (!bitmap->bytes)
2258                 free_bitmap(ctl, bitmap);
2259
2260         return true;
2261 }
2262
2263 /*
2264  * We prefer always to allocate from extent entries, both for clustered and
2265  * non-clustered allocation requests. So when attempting to add a new extent
2266  * entry, try to see if there's adjacent free space in bitmap entries, and if
2267  * there is, migrate that space from the bitmaps to the extent.
2268  * Like this we get better chances of satisfying space allocation requests
2269  * because we attempt to satisfy them based on a single cache entry, and never
2270  * on 2 or more entries - even if the entries represent a contiguous free space
2271  * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2272  * ends).
2273  */
2274 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2275                               struct btrfs_free_space *info,
2276                               bool update_stat)
2277 {
2278         /*
2279          * Only work with disconnected entries, as we can change their offset,
2280          * and must be extent entries.
2281          */
2282         ASSERT(!info->bitmap);
2283         ASSERT(RB_EMPTY_NODE(&info->offset_index));
2284
2285         if (ctl->total_bitmaps > 0) {
2286                 bool stole_end;
2287                 bool stole_front = false;
2288
2289                 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2290                 if (ctl->total_bitmaps > 0)
2291                         stole_front = steal_from_bitmap_to_front(ctl, info,
2292                                                                  update_stat);
2293
2294                 if (stole_end || stole_front)
2295                         try_merge_free_space(ctl, info, update_stat);
2296         }
2297 }
2298
2299 int __btrfs_add_free_space(struct btrfs_fs_info *fs_info,
2300                            struct btrfs_free_space_ctl *ctl,
2301                            u64 offset, u64 bytes)
2302 {
2303         struct btrfs_free_space *info;
2304         int ret = 0;
2305
2306         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2307         if (!info)
2308                 return -ENOMEM;
2309
2310         info->offset = offset;
2311         info->bytes = bytes;
2312         RB_CLEAR_NODE(&info->offset_index);
2313
2314         spin_lock(&ctl->tree_lock);
2315
2316         if (try_merge_free_space(ctl, info, true))
2317                 goto link;
2318
2319         /*
2320          * There was no extent directly to the left or right of this new
2321          * extent then we know we're going to have to allocate a new extent, so
2322          * before we do that see if we need to drop this into a bitmap
2323          */
2324         ret = insert_into_bitmap(ctl, info);
2325         if (ret < 0) {
2326                 goto out;
2327         } else if (ret) {
2328                 ret = 0;
2329                 goto out;
2330         }
2331 link:
2332         /*
2333          * Only steal free space from adjacent bitmaps if we're sure we're not
2334          * going to add the new free space to existing bitmap entries - because
2335          * that would mean unnecessary work that would be reverted. Therefore
2336          * attempt to steal space from bitmaps if we're adding an extent entry.
2337          */
2338         steal_from_bitmap(ctl, info, true);
2339
2340         ret = link_free_space(ctl, info);
2341         if (ret)
2342                 kmem_cache_free(btrfs_free_space_cachep, info);
2343 out:
2344         spin_unlock(&ctl->tree_lock);
2345
2346         if (ret) {
2347                 btrfs_crit(fs_info, "unable to add free space :%d", ret);
2348                 ASSERT(ret != -EEXIST);
2349         }
2350
2351         return ret;
2352 }
2353
2354 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
2355                             u64 offset, u64 bytes)
2356 {
2357         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2358         struct btrfs_free_space *info;
2359         int ret;
2360         bool re_search = false;
2361
2362         spin_lock(&ctl->tree_lock);
2363
2364 again:
2365         ret = 0;
2366         if (!bytes)
2367                 goto out_lock;
2368
2369         info = tree_search_offset(ctl, offset, 0, 0);
2370         if (!info) {
2371                 /*
2372                  * oops didn't find an extent that matched the space we wanted
2373                  * to remove, look for a bitmap instead
2374                  */
2375                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2376                                           1, 0);
2377                 if (!info) {
2378                         /*
2379                          * If we found a partial bit of our free space in a
2380                          * bitmap but then couldn't find the other part this may
2381                          * be a problem, so WARN about it.
2382                          */
2383                         WARN_ON(re_search);
2384                         goto out_lock;
2385                 }
2386         }
2387
2388         re_search = false;
2389         if (!info->bitmap) {
2390                 unlink_free_space(ctl, info);
2391                 if (offset == info->offset) {
2392                         u64 to_free = min(bytes, info->bytes);
2393
2394                         info->bytes -= to_free;
2395                         info->offset += to_free;
2396                         if (info->bytes) {
2397                                 ret = link_free_space(ctl, info);
2398                                 WARN_ON(ret);
2399                         } else {
2400                                 kmem_cache_free(btrfs_free_space_cachep, info);
2401                         }
2402
2403                         offset += to_free;
2404                         bytes -= to_free;
2405                         goto again;
2406                 } else {
2407                         u64 old_end = info->bytes + info->offset;
2408
2409                         info->bytes = offset - info->offset;
2410                         ret = link_free_space(ctl, info);
2411                         WARN_ON(ret);
2412                         if (ret)
2413                                 goto out_lock;
2414
2415                         /* Not enough bytes in this entry to satisfy us */
2416                         if (old_end < offset + bytes) {
2417                                 bytes -= old_end - offset;
2418                                 offset = old_end;
2419                                 goto again;
2420                         } else if (old_end == offset + bytes) {
2421                                 /* all done */
2422                                 goto out_lock;
2423                         }
2424                         spin_unlock(&ctl->tree_lock);
2425
2426                         ret = btrfs_add_free_space(block_group, offset + bytes,
2427                                                    old_end - (offset + bytes));
2428                         WARN_ON(ret);
2429                         goto out;
2430                 }
2431         }
2432
2433         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2434         if (ret == -EAGAIN) {
2435                 re_search = true;
2436                 goto again;
2437         }
2438 out_lock:
2439         spin_unlock(&ctl->tree_lock);
2440 out:
2441         return ret;
2442 }
2443
2444 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
2445                            u64 bytes)
2446 {
2447         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2448         struct btrfs_free_space *info;
2449         struct rb_node *n;
2450         int count = 0;
2451
2452         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2453                 info = rb_entry(n, struct btrfs_free_space, offset_index);
2454                 if (info->bytes >= bytes && !block_group->ro)
2455                         count++;
2456                 btrfs_crit(block_group->fs_info,
2457                            "entry offset %llu, bytes %llu, bitmap %s",
2458                            info->offset, info->bytes,
2459                        (info->bitmap) ? "yes" : "no");
2460         }
2461         btrfs_info(block_group->fs_info, "block group has cluster?: %s",
2462                list_empty(&block_group->cluster_list) ? "no" : "yes");
2463         btrfs_info(block_group->fs_info,
2464                    "%d blocks of free space at or bigger than bytes is", count);
2465 }
2466
2467 void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
2468 {
2469         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2470
2471         spin_lock_init(&ctl->tree_lock);
2472         ctl->unit = block_group->sectorsize;
2473         ctl->start = block_group->key.objectid;
2474         ctl->private = block_group;
2475         ctl->op = &free_space_op;
2476         INIT_LIST_HEAD(&ctl->trimming_ranges);
2477         mutex_init(&ctl->cache_writeout_mutex);
2478
2479         /*
2480          * we only want to have 32k of ram per block group for keeping
2481          * track of free space, and if we pass 1/2 of that we want to
2482          * start converting things over to using bitmaps
2483          */
2484         ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2485 }
2486
2487 /*
2488  * for a given cluster, put all of its extents back into the free
2489  * space cache.  If the block group passed doesn't match the block group
2490  * pointed to by the cluster, someone else raced in and freed the
2491  * cluster already.  In that case, we just return without changing anything
2492  */
2493 static int
2494 __btrfs_return_cluster_to_free_space(
2495                              struct btrfs_block_group_cache *block_group,
2496                              struct btrfs_free_cluster *cluster)
2497 {
2498         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2499         struct btrfs_free_space *entry;
2500         struct rb_node *node;
2501
2502         spin_lock(&cluster->lock);
2503         if (cluster->block_group != block_group)
2504                 goto out;
2505
2506         cluster->block_group = NULL;
2507         cluster->window_start = 0;
2508         list_del_init(&cluster->block_group_list);
2509
2510         node = rb_first(&cluster->root);
2511         while (node) {
2512                 bool bitmap;
2513
2514                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2515                 node = rb_next(&entry->offset_index);
2516                 rb_erase(&entry->offset_index, &cluster->root);
2517                 RB_CLEAR_NODE(&entry->offset_index);
2518
2519                 bitmap = (entry->bitmap != NULL);
2520                 if (!bitmap) {
2521                         try_merge_free_space(ctl, entry, false);
2522                         steal_from_bitmap(ctl, entry, false);
2523                 }
2524                 tree_insert_offset(&ctl->free_space_offset,
2525                                    entry->offset, &entry->offset_index, bitmap);
2526         }
2527         cluster->root = RB_ROOT;
2528
2529 out:
2530         spin_unlock(&cluster->lock);
2531         btrfs_put_block_group(block_group);
2532         return 0;
2533 }
2534
2535 static void __btrfs_remove_free_space_cache_locked(
2536                                 struct btrfs_free_space_ctl *ctl)
2537 {
2538         struct btrfs_free_space *info;
2539         struct rb_node *node;
2540
2541         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2542                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2543                 if (!info->bitmap) {
2544                         unlink_free_space(ctl, info);
2545                         kmem_cache_free(btrfs_free_space_cachep, info);
2546                 } else {
2547                         free_bitmap(ctl, info);
2548                 }
2549
2550                 cond_resched_lock(&ctl->tree_lock);
2551         }
2552 }
2553
2554 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2555 {
2556         spin_lock(&ctl->tree_lock);
2557         __btrfs_remove_free_space_cache_locked(ctl);
2558         spin_unlock(&ctl->tree_lock);
2559 }
2560
2561 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
2562 {
2563         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2564         struct btrfs_free_cluster *cluster;
2565         struct list_head *head;
2566
2567         spin_lock(&ctl->tree_lock);
2568         while ((head = block_group->cluster_list.next) !=
2569                &block_group->cluster_list) {
2570                 cluster = list_entry(head, struct btrfs_free_cluster,
2571                                      block_group_list);
2572
2573                 WARN_ON(cluster->block_group != block_group);
2574                 __btrfs_return_cluster_to_free_space(block_group, cluster);
2575
2576                 cond_resched_lock(&ctl->tree_lock);
2577         }
2578         __btrfs_remove_free_space_cache_locked(ctl);
2579         spin_unlock(&ctl->tree_lock);
2580
2581 }
2582
2583 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
2584                                u64 offset, u64 bytes, u64 empty_size,
2585                                u64 *max_extent_size)
2586 {
2587         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2588         struct btrfs_free_space *entry = NULL;
2589         u64 bytes_search = bytes + empty_size;
2590         u64 ret = 0;
2591         u64 align_gap = 0;
2592         u64 align_gap_len = 0;
2593
2594         spin_lock(&ctl->tree_lock);
2595         entry = find_free_space(ctl, &offset, &bytes_search,
2596                                 block_group->full_stripe_len, max_extent_size);
2597         if (!entry)
2598                 goto out;
2599
2600         ret = offset;
2601         if (entry->bitmap) {
2602                 bitmap_clear_bits(ctl, entry, offset, bytes);
2603                 if (!entry->bytes)
2604                         free_bitmap(ctl, entry);
2605         } else {
2606                 unlink_free_space(ctl, entry);
2607                 align_gap_len = offset - entry->offset;
2608                 align_gap = entry->offset;
2609
2610                 entry->offset = offset + bytes;
2611                 WARN_ON(entry->bytes < bytes + align_gap_len);
2612
2613                 entry->bytes -= bytes + align_gap_len;
2614                 if (!entry->bytes)
2615                         kmem_cache_free(btrfs_free_space_cachep, entry);
2616                 else
2617                         link_free_space(ctl, entry);
2618         }
2619 out:
2620         spin_unlock(&ctl->tree_lock);
2621
2622         if (align_gap_len)
2623                 __btrfs_add_free_space(block_group->fs_info, ctl,
2624                                        align_gap, align_gap_len);
2625         return ret;
2626 }
2627
2628 /*
2629  * given a cluster, put all of its extents back into the free space
2630  * cache.  If a block group is passed, this function will only free
2631  * a cluster that belongs to the passed block group.
2632  *
2633  * Otherwise, it'll get a reference on the block group pointed to by the
2634  * cluster and remove the cluster from it.
2635  */
2636 int btrfs_return_cluster_to_free_space(
2637                                struct btrfs_block_group_cache *block_group,
2638                                struct btrfs_free_cluster *cluster)
2639 {
2640         struct btrfs_free_space_ctl *ctl;
2641         int ret;
2642
2643         /* first, get a safe pointer to the block group */
2644         spin_lock(&cluster->lock);
2645         if (!block_group) {
2646                 block_group = cluster->block_group;
2647                 if (!block_group) {
2648                         spin_unlock(&cluster->lock);
2649                         return 0;
2650                 }
2651         } else if (cluster->block_group != block_group) {
2652                 /* someone else has already freed it don't redo their work */
2653                 spin_unlock(&cluster->lock);
2654                 return 0;
2655         }
2656         atomic_inc(&block_group->count);
2657         spin_unlock(&cluster->lock);
2658
2659         ctl = block_group->free_space_ctl;
2660
2661         /* now return any extents the cluster had on it */
2662         spin_lock(&ctl->tree_lock);
2663         ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2664         spin_unlock(&ctl->tree_lock);
2665
2666         /* finally drop our ref */
2667         btrfs_put_block_group(block_group);
2668         return ret;
2669 }
2670
2671 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
2672                                    struct btrfs_free_cluster *cluster,
2673                                    struct btrfs_free_space *entry,
2674                                    u64 bytes, u64 min_start,
2675                                    u64 *max_extent_size)
2676 {
2677         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2678         int err;
2679         u64 search_start = cluster->window_start;
2680         u64 search_bytes = bytes;
2681         u64 ret = 0;
2682
2683         search_start = min_start;
2684         search_bytes = bytes;
2685
2686         err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
2687         if (err) {
2688                 if (search_bytes > *max_extent_size)
2689                         *max_extent_size = search_bytes;
2690                 return 0;
2691         }
2692
2693         ret = search_start;
2694         __bitmap_clear_bits(ctl, entry, ret, bytes);
2695
2696         return ret;
2697 }
2698
2699 /*
2700  * given a cluster, try to allocate 'bytes' from it, returns 0
2701  * if it couldn't find anything suitably large, or a logical disk offset
2702  * if things worked out
2703  */
2704 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
2705                              struct btrfs_free_cluster *cluster, u64 bytes,
2706                              u64 min_start, u64 *max_extent_size)
2707 {
2708         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2709         struct btrfs_free_space *entry = NULL;
2710         struct rb_node *node;
2711         u64 ret = 0;
2712
2713         spin_lock(&cluster->lock);
2714         if (bytes > cluster->max_size)
2715                 goto out;
2716
2717         if (cluster->block_group != block_group)
2718                 goto out;
2719
2720         node = rb_first(&cluster->root);
2721         if (!node)
2722                 goto out;
2723
2724         entry = rb_entry(node, struct btrfs_free_space, offset_index);
2725         while (1) {
2726                 if (entry->bytes < bytes && entry->bytes > *max_extent_size)
2727                         *max_extent_size = entry->bytes;
2728
2729                 if (entry->bytes < bytes ||
2730                     (!entry->bitmap && entry->offset < min_start)) {
2731                         node = rb_next(&entry->offset_index);
2732                         if (!node)
2733                                 break;
2734                         entry = rb_entry(node, struct btrfs_free_space,
2735                                          offset_index);
2736                         continue;
2737                 }
2738
2739                 if (entry->bitmap) {
2740                         ret = btrfs_alloc_from_bitmap(block_group,
2741                                                       cluster, entry, bytes,
2742                                                       cluster->window_start,
2743                                                       max_extent_size);
2744                         if (ret == 0) {
2745                                 node = rb_next(&entry->offset_index);
2746                                 if (!node)
2747                                         break;
2748                                 entry = rb_entry(node, struct btrfs_free_space,
2749                                                  offset_index);
2750                                 continue;
2751                         }
2752                         cluster->window_start += bytes;
2753                 } else {
2754                         ret = entry->offset;
2755
2756                         entry->offset += bytes;
2757                         entry->bytes -= bytes;
2758                 }
2759
2760                 if (entry->bytes == 0)
2761                         rb_erase(&entry->offset_index, &cluster->root);
2762                 break;
2763         }
2764 out:
2765         spin_unlock(&cluster->lock);
2766
2767         if (!ret)
2768                 return 0;
2769
2770         spin_lock(&ctl->tree_lock);
2771
2772         ctl->free_space -= bytes;
2773         if (entry->bytes == 0) {
2774                 ctl->free_extents--;
2775                 if (entry->bitmap) {
2776                         kfree(entry->bitmap);
2777                         ctl->total_bitmaps--;
2778                         ctl->op->recalc_thresholds(ctl);
2779                 }
2780                 kmem_cache_free(btrfs_free_space_cachep, entry);
2781         }
2782
2783         spin_unlock(&ctl->tree_lock);
2784
2785         return ret;
2786 }
2787
2788 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2789                                 struct btrfs_free_space *entry,
2790                                 struct btrfs_free_cluster *cluster,
2791                                 u64 offset, u64 bytes,
2792                                 u64 cont1_bytes, u64 min_bytes)
2793 {
2794         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2795         unsigned long next_zero;
2796         unsigned long i;
2797         unsigned long want_bits;
2798         unsigned long min_bits;
2799         unsigned long found_bits;
2800         unsigned long max_bits = 0;
2801         unsigned long start = 0;
2802         unsigned long total_found = 0;
2803         int ret;
2804
2805         i = offset_to_bit(entry->offset, ctl->unit,
2806                           max_t(u64, offset, entry->offset));
2807         want_bits = bytes_to_bits(bytes, ctl->unit);
2808         min_bits = bytes_to_bits(min_bytes, ctl->unit);
2809
2810         /*
2811          * Don't bother looking for a cluster in this bitmap if it's heavily
2812          * fragmented.
2813          */
2814         if (entry->max_extent_size &&
2815             entry->max_extent_size < cont1_bytes)
2816                 return -ENOSPC;
2817 again:
2818         found_bits = 0;
2819         for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
2820                 next_zero = find_next_zero_bit(entry->bitmap,
2821                                                BITS_PER_BITMAP, i);
2822                 if (next_zero - i >= min_bits) {
2823                         found_bits = next_zero - i;
2824                         if (found_bits > max_bits)
2825                                 max_bits = found_bits;
2826                         break;
2827                 }
2828                 if (next_zero - i > max_bits)
2829                         max_bits = next_zero - i;
2830                 i = next_zero;
2831         }
2832
2833         if (!found_bits) {
2834                 entry->max_extent_size = (u64)max_bits * ctl->unit;
2835                 return -ENOSPC;
2836         }
2837
2838         if (!total_found) {
2839                 start = i;
2840                 cluster->max_size = 0;
2841         }
2842
2843         total_found += found_bits;
2844
2845         if (cluster->max_size < found_bits * ctl->unit)
2846                 cluster->max_size = found_bits * ctl->unit;
2847
2848         if (total_found < want_bits || cluster->max_size < cont1_bytes) {
2849                 i = next_zero + 1;
2850                 goto again;
2851         }
2852
2853         cluster->window_start = start * ctl->unit + entry->offset;
2854         rb_erase(&entry->offset_index, &ctl->free_space_offset);
2855         ret = tree_insert_offset(&cluster->root, entry->offset,
2856                                  &entry->offset_index, 1);
2857         ASSERT(!ret); /* -EEXIST; Logic error */
2858
2859         trace_btrfs_setup_cluster(block_group, cluster,
2860                                   total_found * ctl->unit, 1);
2861         return 0;
2862 }
2863
2864 /*
2865  * This searches the block group for just extents to fill the cluster with.
2866  * Try to find a cluster with at least bytes total bytes, at least one
2867  * extent of cont1_bytes, and other clusters of at least min_bytes.
2868  */
2869 static noinline int
2870 setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2871                         struct btrfs_free_cluster *cluster,
2872                         struct list_head *bitmaps, u64 offset, u64 bytes,
2873                         u64 cont1_bytes, u64 min_bytes)
2874 {
2875         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2876         struct btrfs_free_space *first = NULL;
2877         struct btrfs_free_space *entry = NULL;
2878         struct btrfs_free_space *last;
2879         struct rb_node *node;
2880         u64 window_free;
2881         u64 max_extent;
2882         u64 total_size = 0;
2883
2884         entry = tree_search_offset(ctl, offset, 0, 1);
2885         if (!entry)
2886                 return -ENOSPC;
2887
2888         /*
2889          * We don't want bitmaps, so just move along until we find a normal
2890          * extent entry.
2891          */
2892         while (entry->bitmap || entry->bytes < min_bytes) {
2893                 if (entry->bitmap && list_empty(&entry->list))
2894                         list_add_tail(&entry->list, bitmaps);
2895                 node = rb_next(&entry->offset_index);
2896                 if (!node)
2897                         return -ENOSPC;
2898                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2899         }
2900
2901         window_free = entry->bytes;
2902         max_extent = entry->bytes;
2903         first = entry;
2904         last = entry;
2905
2906         for (node = rb_next(&entry->offset_index); node;
2907              node = rb_next(&entry->offset_index)) {
2908                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2909
2910                 if (entry->bitmap) {
2911                         if (list_empty(&entry->list))
2912                                 list_add_tail(&entry->list, bitmaps);
2913                         continue;
2914                 }
2915
2916                 if (entry->bytes < min_bytes)
2917                         continue;
2918
2919                 last = entry;
2920                 window_free += entry->bytes;
2921                 if (entry->bytes > max_extent)
2922                         max_extent = entry->bytes;
2923         }
2924
2925         if (window_free < bytes || max_extent < cont1_bytes)
2926                 return -ENOSPC;
2927
2928         cluster->window_start = first->offset;
2929
2930         node = &first->offset_index;
2931
2932         /*
2933          * now we've found our entries, pull them out of the free space
2934          * cache and put them into the cluster rbtree
2935          */
2936         do {
2937                 int ret;
2938
2939                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2940                 node = rb_next(&entry->offset_index);
2941                 if (entry->bitmap || entry->bytes < min_bytes)
2942                         continue;
2943
2944                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2945                 ret = tree_insert_offset(&cluster->root, entry->offset,
2946                                          &entry->offset_index, 0);
2947                 total_size += entry->bytes;
2948                 ASSERT(!ret); /* -EEXIST; Logic error */
2949         } while (node && entry != last);
2950
2951         cluster->max_size = max_extent;
2952         trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
2953         return 0;
2954 }
2955
2956 /*
2957  * This specifically looks for bitmaps that may work in the cluster, we assume
2958  * that we have already failed to find extents that will work.
2959  */
2960 static noinline int
2961 setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2962                      struct btrfs_free_cluster *cluster,
2963                      struct list_head *bitmaps, u64 offset, u64 bytes,
2964                      u64 cont1_bytes, u64 min_bytes)
2965 {
2966         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2967         struct btrfs_free_space *entry = NULL;
2968         int ret = -ENOSPC;
2969         u64 bitmap_offset = offset_to_bitmap(ctl, offset);
2970
2971         if (ctl->total_bitmaps == 0)
2972                 return -ENOSPC;
2973
2974         /*
2975          * The bitmap that covers offset won't be in the list unless offset
2976          * is just its start offset.
2977          */
2978         if (!list_empty(bitmaps))
2979                 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
2980
2981         if (!entry || entry->offset != bitmap_offset) {
2982                 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
2983                 if (entry && list_empty(&entry->list))
2984                         list_add(&entry->list, bitmaps);
2985         }
2986
2987         list_for_each_entry(entry, bitmaps, list) {
2988                 if (entry->bytes < bytes)
2989                         continue;
2990                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2991                                            bytes, cont1_bytes, min_bytes);
2992                 if (!ret)
2993                         return 0;
2994         }
2995
2996         /*
2997          * The bitmaps list has all the bitmaps that record free space
2998          * starting after offset, so no more search is required.
2999          */
3000         return -ENOSPC;
3001 }
3002
3003 /*
3004  * here we try to find a cluster of blocks in a block group.  The goal
3005  * is to find at least bytes+empty_size.
3006  * We might not find them all in one contiguous area.
3007  *
3008  * returns zero and sets up cluster if things worked out, otherwise
3009  * it returns -enospc
3010  */
3011 int btrfs_find_space_cluster(struct btrfs_root *root,
3012                              struct btrfs_block_group_cache *block_group,
3013                              struct btrfs_free_cluster *cluster,
3014                              u64 offset, u64 bytes, u64 empty_size)
3015 {
3016         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3017         struct btrfs_free_space *entry, *tmp;
3018         LIST_HEAD(bitmaps);
3019         u64 min_bytes;
3020         u64 cont1_bytes;
3021         int ret;
3022
3023         /*
3024          * Choose the minimum extent size we'll require for this
3025          * cluster.  For SSD_SPREAD, don't allow any fragmentation.
3026          * For metadata, allow allocates with smaller extents.  For
3027          * data, keep it dense.
3028          */
3029         if (btrfs_test_opt(root->fs_info, SSD_SPREAD)) {
3030                 cont1_bytes = min_bytes = bytes + empty_size;
3031         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3032                 cont1_bytes = bytes;
3033                 min_bytes = block_group->sectorsize;
3034         } else {
3035                 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3036                 min_bytes = block_group->sectorsize;
3037         }
3038
3039         spin_lock(&ctl->tree_lock);
3040
3041         /*
3042          * If we know we don't have enough space to make a cluster don't even
3043          * bother doing all the work to try and find one.
3044          */
3045         if (ctl->free_space < bytes) {
3046                 spin_unlock(&ctl->tree_lock);
3047                 return -ENOSPC;
3048         }
3049
3050         spin_lock(&cluster->lock);
3051
3052         /* someone already found a cluster, hooray */
3053         if (cluster->block_group) {
3054                 ret = 0;
3055                 goto out;
3056         }
3057
3058         trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3059                                  min_bytes);
3060
3061         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3062                                       bytes + empty_size,
3063                                       cont1_bytes, min_bytes);
3064         if (ret)
3065                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3066                                            offset, bytes + empty_size,
3067                                            cont1_bytes, min_bytes);
3068
3069         /* Clear our temporary list */
3070         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3071                 list_del_init(&entry->list);
3072
3073         if (!ret) {
3074                 atomic_inc(&block_group->count);
3075                 list_add_tail(&cluster->block_group_list,
3076                               &block_group->cluster_list);
3077                 cluster->block_group = block_group;
3078         } else {
3079                 trace_btrfs_failed_cluster_setup(block_group);
3080         }
3081 out:
3082         spin_unlock(&cluster->lock);
3083         spin_unlock(&ctl->tree_lock);
3084
3085         return ret;
3086 }
3087
3088 /*
3089  * simple code to zero out a cluster
3090  */
3091 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3092 {
3093         spin_lock_init(&cluster->lock);
3094         spin_lock_init(&cluster->refill_lock);
3095         cluster->root = RB_ROOT;
3096         cluster->max_size = 0;
3097         cluster->fragmented = false;
3098         INIT_LIST_HEAD(&cluster->block_group_list);
3099         cluster->block_group = NULL;
3100 }
3101
3102 static int do_trimming(struct btrfs_block_group_cache *block_group,
3103                        u64 *total_trimmed, u64 start, u64 bytes,
3104                        u64 reserved_start, u64 reserved_bytes,
3105                        struct btrfs_trim_range *trim_entry)
3106 {
3107         struct btrfs_space_info *space_info = block_group->space_info;
3108         struct btrfs_fs_info *fs_info = block_group->fs_info;
3109         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3110         int ret;
3111         int update = 0;
3112         u64 trimmed = 0;
3113
3114         spin_lock(&space_info->lock);
3115         spin_lock(&block_group->lock);
3116         if (!block_group->ro) {
3117                 block_group->reserved += reserved_bytes;
3118                 space_info->bytes_reserved += reserved_bytes;
3119                 update = 1;
3120         }
3121         spin_unlock(&block_group->lock);
3122         spin_unlock(&space_info->lock);
3123
3124         ret = btrfs_discard_extent(fs_info->extent_root,
3125                                    start, bytes, &trimmed);
3126         if (!ret)
3127                 *total_trimmed += trimmed;
3128
3129         mutex_lock(&ctl->cache_writeout_mutex);
3130         btrfs_add_free_space(block_group, reserved_start, reserved_bytes);
3131         list_del(&trim_entry->list);
3132         mutex_unlock(&ctl->cache_writeout_mutex);
3133
3134         if (update) {
3135                 spin_lock(&space_info->lock);
3136                 spin_lock(&block_group->lock);
3137                 if (block_group->ro)
3138                         space_info->bytes_readonly += reserved_bytes;
3139                 block_group->reserved -= reserved_bytes;
3140                 space_info->bytes_reserved -= reserved_bytes;
3141                 spin_unlock(&space_info->lock);
3142                 spin_unlock(&block_group->lock);
3143         }
3144
3145         return ret;
3146 }
3147
3148 static int trim_no_bitmap(struct btrfs_block_group_cache *block_group,
3149                           u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3150 {
3151         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3152         struct btrfs_free_space *entry;
3153         struct rb_node *node;
3154         int ret = 0;
3155         u64 extent_start;
3156         u64 extent_bytes;
3157         u64 bytes;
3158
3159         while (start < end) {
3160                 struct btrfs_trim_range trim_entry;
3161
3162                 mutex_lock(&ctl->cache_writeout_mutex);
3163                 spin_lock(&ctl->tree_lock);
3164
3165                 if (ctl->free_space < minlen) {
3166                         spin_unlock(&ctl->tree_lock);
3167                         mutex_unlock(&ctl->cache_writeout_mutex);
3168                         break;
3169                 }
3170
3171                 entry = tree_search_offset(ctl, start, 0, 1);
3172                 if (!entry) {
3173                         spin_unlock(&ctl->tree_lock);
3174                         mutex_unlock(&ctl->cache_writeout_mutex);
3175                         break;
3176                 }
3177
3178                 /* skip bitmaps */
3179                 while (entry->bitmap) {
3180                         node = rb_next(&entry->offset_index);
3181                         if (!node) {
3182                                 spin_unlock(&ctl->tree_lock);
3183                                 mutex_unlock(&ctl->cache_writeout_mutex);
3184                                 goto out;
3185                         }
3186                         entry = rb_entry(node, struct btrfs_free_space,
3187                                          offset_index);
3188                 }
3189
3190                 if (entry->offset >= end) {
3191                         spin_unlock(&ctl->tree_lock);
3192                         mutex_unlock(&ctl->cache_writeout_mutex);
3193                         break;
3194                 }
3195
3196                 extent_start = entry->offset;
3197                 extent_bytes = entry->bytes;
3198                 start = max(start, extent_start);
3199                 bytes = min(extent_start + extent_bytes, end) - start;
3200                 if (bytes < minlen) {
3201                         spin_unlock(&ctl->tree_lock);
3202                         mutex_unlock(&ctl->cache_writeout_mutex);
3203                         goto next;
3204                 }
3205
3206                 unlink_free_space(ctl, entry);
3207                 kmem_cache_free(btrfs_free_space_cachep, entry);
3208
3209                 spin_unlock(&ctl->tree_lock);
3210                 trim_entry.start = extent_start;
3211                 trim_entry.bytes = extent_bytes;
3212                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3213                 mutex_unlock(&ctl->cache_writeout_mutex);
3214
3215                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3216                                   extent_start, extent_bytes, &trim_entry);
3217                 if (ret)
3218                         break;
3219 next:
3220                 start += bytes;
3221
3222                 if (fatal_signal_pending(current)) {
3223                         ret = -ERESTARTSYS;
3224                         break;
3225                 }
3226
3227                 cond_resched();
3228         }
3229 out:
3230         return ret;
3231 }
3232
3233 static int trim_bitmaps(struct btrfs_block_group_cache *block_group,
3234                         u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3235 {
3236         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3237         struct btrfs_free_space *entry;
3238         int ret = 0;
3239         int ret2;
3240         u64 bytes;
3241         u64 offset = offset_to_bitmap(ctl, start);
3242
3243         while (offset < end) {
3244                 bool next_bitmap = false;
3245                 struct btrfs_trim_range trim_entry;
3246
3247                 mutex_lock(&ctl->cache_writeout_mutex);
3248                 spin_lock(&ctl->tree_lock);
3249
3250                 if (ctl->free_space < minlen) {
3251                         spin_unlock(&ctl->tree_lock);
3252                         mutex_unlock(&ctl->cache_writeout_mutex);
3253                         break;
3254                 }
3255
3256                 entry = tree_search_offset(ctl, offset, 1, 0);
3257                 if (!entry) {
3258                         spin_unlock(&ctl->tree_lock);
3259                         mutex_unlock(&ctl->cache_writeout_mutex);
3260                         next_bitmap = true;
3261                         goto next;
3262                 }
3263
3264                 bytes = minlen;
3265                 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3266                 if (ret2 || start >= end) {
3267                         spin_unlock(&ctl->tree_lock);
3268                         mutex_unlock(&ctl->cache_writeout_mutex);
3269                         next_bitmap = true;
3270                         goto next;
3271                 }
3272
3273                 bytes = min(bytes, end - start);
3274                 if (bytes < minlen) {
3275                         spin_unlock(&ctl->tree_lock);
3276                         mutex_unlock(&ctl->cache_writeout_mutex);
3277                         goto next;
3278                 }
3279
3280                 bitmap_clear_bits(ctl, entry, start, bytes);
3281                 if (entry->bytes == 0)
3282                         free_bitmap(ctl, entry);
3283
3284                 spin_unlock(&ctl->tree_lock);
3285                 trim_entry.start = start;
3286                 trim_entry.bytes = bytes;
3287                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3288                 mutex_unlock(&ctl->cache_writeout_mutex);
3289
3290                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3291                                   start, bytes, &trim_entry);
3292                 if (ret)
3293                         break;
3294 next:
3295                 if (next_bitmap) {
3296                         offset += BITS_PER_BITMAP * ctl->unit;
3297                 } else {
3298                         start += bytes;
3299                         if (start >= offset + BITS_PER_BITMAP * ctl->unit)
3300                                 offset += BITS_PER_BITMAP * ctl->unit;
3301                 }
3302
3303                 if (fatal_signal_pending(current)) {
3304                         ret = -ERESTARTSYS;
3305                         break;
3306                 }
3307
3308                 cond_resched();
3309         }
3310
3311         return ret;
3312 }
3313
3314 void btrfs_get_block_group_trimming(struct btrfs_block_group_cache *cache)
3315 {
3316         atomic_inc(&cache->trimming);
3317 }
3318
3319 void btrfs_put_block_group_trimming(struct btrfs_block_group_cache *block_group)
3320 {
3321         struct extent_map_tree *em_tree;
3322         struct extent_map *em;
3323         bool cleanup;
3324
3325         spin_lock(&block_group->lock);
3326         cleanup = (atomic_dec_and_test(&block_group->trimming) &&
3327                    block_group->removed);
3328         spin_unlock(&block_group->lock);
3329
3330         if (cleanup) {
3331                 lock_chunks(block_group->fs_info->chunk_root);
3332                 em_tree = &block_group->fs_info->mapping_tree.map_tree;
3333                 write_lock(&em_tree->lock);
3334                 em = lookup_extent_mapping(em_tree, block_group->key.objectid,
3335                                            1);
3336                 BUG_ON(!em); /* logic error, can't happen */
3337                 /*
3338                  * remove_extent_mapping() will delete us from the pinned_chunks
3339                  * list, which is protected by the chunk mutex.
3340                  */
3341                 remove_extent_mapping(em_tree, em);
3342                 write_unlock(&em_tree->lock);
3343                 unlock_chunks(block_group->fs_info->chunk_root);
3344
3345                 /* once for us and once for the tree */
3346                 free_extent_map(em);
3347                 free_extent_map(em);
3348
3349                 /*
3350                  * We've left one free space entry and other tasks trimming
3351                  * this block group have left 1 entry each one. Free them.
3352                  */
3353                 __btrfs_remove_free_space_cache(block_group->free_space_ctl);
3354         }
3355 }
3356
3357 int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
3358                            u64 *trimmed, u64 start, u64 end, u64 minlen)
3359 {
3360         int ret;
3361
3362         *trimmed = 0;
3363
3364         spin_lock(&block_group->lock);
3365         if (block_group->removed) {
3366                 spin_unlock(&block_group->lock);
3367                 return 0;
3368         }
3369         btrfs_get_block_group_trimming(block_group);
3370         spin_unlock(&block_group->lock);
3371
3372         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen);
3373         if (ret)
3374                 goto out;
3375
3376         ret = trim_bitmaps(block_group, trimmed, start, end, minlen);
3377 out:
3378         btrfs_put_block_group_trimming(block_group);
3379         return ret;
3380 }
3381
3382 /*
3383  * Find the left-most item in the cache tree, and then return the
3384  * smallest inode number in the item.
3385  *
3386  * Note: the returned inode number may not be the smallest one in
3387  * the tree, if the left-most item is a bitmap.
3388  */
3389 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
3390 {
3391         struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
3392         struct btrfs_free_space *entry = NULL;
3393         u64 ino = 0;
3394
3395         spin_lock(&ctl->tree_lock);
3396
3397         if (RB_EMPTY_ROOT(&ctl->free_space_offset))
3398                 goto out;
3399
3400         entry = rb_entry(rb_first(&ctl->free_space_offset),
3401                          struct btrfs_free_space, offset_index);
3402
3403         if (!entry->bitmap) {
3404                 ino = entry->offset;
3405
3406                 unlink_free_space(ctl, entry);
3407                 entry->offset++;
3408                 entry->bytes--;
3409                 if (!entry->bytes)
3410                         kmem_cache_free(btrfs_free_space_cachep, entry);
3411                 else
3412                         link_free_space(ctl, entry);
3413         } else {
3414                 u64 offset = 0;
3415                 u64 count = 1;
3416                 int ret;
3417
3418                 ret = search_bitmap(ctl, entry, &offset, &count, true);
3419                 /* Logic error; Should be empty if it can't find anything */
3420                 ASSERT(!ret);
3421
3422                 ino = offset;
3423                 bitmap_clear_bits(ctl, entry, offset, 1);
3424                 if (entry->bytes == 0)
3425                         free_bitmap(ctl, entry);
3426         }
3427 out:
3428         spin_unlock(&ctl->tree_lock);
3429
3430         return ino;
3431 }
3432
3433 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
3434                                     struct btrfs_path *path)
3435 {
3436         struct inode *inode = NULL;
3437
3438         spin_lock(&root->ino_cache_lock);
3439         if (root->ino_cache_inode)
3440                 inode = igrab(root->ino_cache_inode);
3441         spin_unlock(&root->ino_cache_lock);
3442         if (inode)
3443                 return inode;
3444
3445         inode = __lookup_free_space_inode(root, path, 0);
3446         if (IS_ERR(inode))
3447                 return inode;
3448
3449         spin_lock(&root->ino_cache_lock);
3450         if (!btrfs_fs_closing(root->fs_info))
3451                 root->ino_cache_inode = igrab(inode);
3452         spin_unlock(&root->ino_cache_lock);
3453
3454         return inode;
3455 }
3456
3457 int create_free_ino_inode(struct btrfs_root *root,
3458                           struct btrfs_trans_handle *trans,
3459                           struct btrfs_path *path)
3460 {
3461         return __create_free_space_inode(root, trans, path,
3462                                          BTRFS_FREE_INO_OBJECTID, 0);
3463 }
3464
3465 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
3466 {
3467         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3468         struct btrfs_path *path;
3469         struct inode *inode;
3470         int ret = 0;
3471         u64 root_gen = btrfs_root_generation(&root->root_item);
3472
3473         if (!btrfs_test_opt(root->fs_info, INODE_MAP_CACHE))
3474                 return 0;
3475
3476         /*
3477          * If we're unmounting then just return, since this does a search on the
3478          * normal root and not the commit root and we could deadlock.
3479          */
3480         if (btrfs_fs_closing(fs_info))
3481                 return 0;
3482
3483         path = btrfs_alloc_path();
3484         if (!path)
3485                 return 0;
3486
3487         inode = lookup_free_ino_inode(root, path);
3488         if (IS_ERR(inode))
3489                 goto out;
3490
3491         if (root_gen != BTRFS_I(inode)->generation)
3492                 goto out_put;
3493
3494         ret = __load_free_space_cache(root, inode, ctl, path, 0);
3495
3496         if (ret < 0)
3497                 btrfs_err(fs_info,
3498                         "failed to load free ino cache for root %llu",
3499                         root->root_key.objectid);
3500 out_put:
3501         iput(inode);
3502 out:
3503         btrfs_free_path(path);
3504         return ret;
3505 }
3506
3507 int btrfs_write_out_ino_cache(struct btrfs_root *root,
3508                               struct btrfs_trans_handle *trans,
3509                               struct btrfs_path *path,
3510                               struct inode *inode)
3511 {
3512         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3513         int ret;
3514         struct btrfs_io_ctl io_ctl;
3515         bool release_metadata = true;
3516
3517         if (!btrfs_test_opt(root->fs_info, INODE_MAP_CACHE))
3518                 return 0;
3519
3520         memset(&io_ctl, 0, sizeof(io_ctl));
3521         ret = __btrfs_write_out_cache(root, inode, ctl, NULL, &io_ctl,
3522                                       trans, path, 0);
3523         if (!ret) {
3524                 /*
3525                  * At this point writepages() didn't error out, so our metadata
3526                  * reservation is released when the writeback finishes, at
3527                  * inode.c:btrfs_finish_ordered_io(), regardless of it finishing
3528                  * with or without an error.
3529                  */
3530                 release_metadata = false;
3531                 ret = btrfs_wait_cache_io(root, trans, NULL, &io_ctl, path, 0);
3532         }
3533
3534         if (ret) {
3535                 if (release_metadata)
3536                         btrfs_delalloc_release_metadata(inode, inode->i_size);
3537 #ifdef DEBUG
3538                 btrfs_err(root->fs_info,
3539                         "failed to write free ino cache for root %llu",
3540                         root->root_key.objectid);
3541 #endif
3542         }
3543
3544         return ret;
3545 }
3546
3547 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3548 /*
3549  * Use this if you need to make a bitmap or extent entry specifically, it
3550  * doesn't do any of the merging that add_free_space does, this acts a lot like
3551  * how the free space cache loading stuff works, so you can get really weird
3552  * configurations.
3553  */
3554 int test_add_free_space_entry(struct btrfs_block_group_cache *cache,
3555                               u64 offset, u64 bytes, bool bitmap)
3556 {
3557         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3558         struct btrfs_free_space *info = NULL, *bitmap_info;
3559         void *map = NULL;
3560         u64 bytes_added;
3561         int ret;
3562
3563 again:
3564         if (!info) {
3565                 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
3566                 if (!info)
3567                         return -ENOMEM;
3568         }
3569
3570         if (!bitmap) {
3571                 spin_lock(&ctl->tree_lock);
3572                 info->offset = offset;
3573                 info->bytes = bytes;
3574                 info->max_extent_size = 0;
3575                 ret = link_free_space(ctl, info);
3576                 spin_unlock(&ctl->tree_lock);
3577                 if (ret)
3578                         kmem_cache_free(btrfs_free_space_cachep, info);
3579                 return ret;
3580         }
3581
3582         if (!map) {
3583                 map = kzalloc(PAGE_SIZE, GFP_NOFS);
3584                 if (!map) {
3585                         kmem_cache_free(btrfs_free_space_cachep, info);
3586                         return -ENOMEM;
3587                 }
3588         }
3589
3590         spin_lock(&ctl->tree_lock);
3591         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3592                                          1, 0);
3593         if (!bitmap_info) {
3594                 info->bitmap = map;
3595                 map = NULL;
3596                 add_new_bitmap(ctl, info, offset);
3597                 bitmap_info = info;
3598                 info = NULL;
3599         }
3600
3601         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
3602
3603         bytes -= bytes_added;
3604         offset += bytes_added;
3605         spin_unlock(&ctl->tree_lock);
3606
3607         if (bytes)
3608                 goto again;
3609
3610         if (info)
3611                 kmem_cache_free(btrfs_free_space_cachep, info);
3612         if (map)
3613                 kfree(map);
3614         return 0;
3615 }
3616
3617 /*
3618  * Checks to see if the given range is in the free space cache.  This is really
3619  * just used to check the absence of space, so if there is free space in the
3620  * range at all we will return 1.
3621  */
3622 int test_check_exists(struct btrfs_block_group_cache *cache,
3623                       u64 offset, u64 bytes)
3624 {
3625         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3626         struct btrfs_free_space *info;
3627         int ret = 0;
3628
3629         spin_lock(&ctl->tree_lock);
3630         info = tree_search_offset(ctl, offset, 0, 0);
3631         if (!info) {
3632                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3633                                           1, 0);
3634                 if (!info)
3635                         goto out;
3636         }
3637
3638 have_info:
3639         if (info->bitmap) {
3640                 u64 bit_off, bit_bytes;
3641                 struct rb_node *n;
3642                 struct btrfs_free_space *tmp;
3643
3644                 bit_off = offset;
3645                 bit_bytes = ctl->unit;
3646                 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
3647                 if (!ret) {
3648                         if (bit_off == offset) {
3649                                 ret = 1;
3650                                 goto out;
3651                         } else if (bit_off > offset &&
3652                                    offset + bytes > bit_off) {
3653                                 ret = 1;
3654                                 goto out;
3655                         }
3656                 }
3657
3658                 n = rb_prev(&info->offset_index);
3659                 while (n) {
3660                         tmp = rb_entry(n, struct btrfs_free_space,
3661                                        offset_index);
3662                         if (tmp->offset + tmp->bytes < offset)
3663                                 break;
3664                         if (offset + bytes < tmp->offset) {
3665                                 n = rb_prev(&tmp->offset_index);
3666                                 continue;
3667                         }
3668                         info = tmp;
3669                         goto have_info;
3670                 }
3671
3672                 n = rb_next(&info->offset_index);
3673                 while (n) {
3674                         tmp = rb_entry(n, struct btrfs_free_space,
3675                                        offset_index);
3676                         if (offset + bytes < tmp->offset)
3677                                 break;
3678                         if (tmp->offset + tmp->bytes < offset) {
3679                                 n = rb_next(&tmp->offset_index);
3680                                 continue;
3681                         }
3682                         info = tmp;
3683                         goto have_info;
3684                 }
3685
3686                 ret = 0;
3687                 goto out;
3688         }
3689
3690         if (info->offset == offset) {
3691                 ret = 1;
3692                 goto out;
3693         }
3694
3695         if (offset > info->offset && offset < info->offset + info->bytes)
3696                 ret = 1;
3697 out:
3698         spin_unlock(&ctl->tree_lock);
3699         return ret;
3700 }
3701 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */